Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+TTS-Spaces-Arena-25-Dec-2024.png filter=lfs diff=lfs merge=lfs -text
+HEARME.wav filter=lfs diff=lfs merge=lfs -text
+demo/af_sky.wav filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,161 @@
+---
+license: apache-2.0
+language:
+- en
+base_model:
+- yl4579/StyleTTS2-LJSpeech
+pipeline_tag: text-to-speech
+---
+❤️ Kokoro Discord Server: https://discord.gg/QuGxSWBfQy
+
+📣 Got Synthetic Data? Want Trained Voicepacks? See https://hf.co/posts/hexgrad/418806998707773
+
+<audio controls><source src="https://huggingface.co/hexgrad/Kokoro-82M/resolve/main/demo/HEARME.wav" type="audio/wav"></audio>
+
+**Kokoro** is a frontier TTS model for its size of **82 million parameters** (text in/audio out).
+
+On 25 Dec 2024, Kokoro v0.19 weights were permissively released in full fp32 precision under an Apache 2.0 license. As of 2 Jan 2025, 10 unique Voicepacks have been released, and a `.onnx` version of v0.19 is available.
+
+In the weeks leading up to its release, Kokoro v0.19 was the #1🥇 ranked model in [TTS Spaces Arena](https://huggingface.co/hexgrad/Kokoro-82M#evaluation). Kokoro had achieved higher Elo in this single-voice Arena setting over other models, using fewer parameters and less data:
+1. **Kokoro v0.19: 82M params, Apache, trained on <100 hours of audio**
+2. XTTS v2: 467M, CPML, >10k hours
+3. Edge TTS: Microsoft, proprietary
+4. MetaVoice: 1.2B, Apache, 100k hours
+5. Parler Mini: 880M, Apache, 45k hours
+6. Fish Speech: ~500M, CC-BY-NC-SA, 1M hours
+
+Kokoro's ability to top this Elo ladder suggests that the scaling law (Elo vs compute/data/params) for traditional TTS models might have a steeper slope than previously expected.
+
+You can find a hosted demo at [hf.co/spaces/hexgrad/Kokoro-TTS](https://huggingface.co/spaces/hexgrad/Kokoro-TTS).
+
+### Usage
+
+The following can be run in a single cell on [Google Colab](https://colab.research.google.com/).
+```py
+# 1️⃣ Install dependencies silently
+!git lfs install
+!git clone https://huggingface.co/hexgrad/Kokoro-82M
+%cd Kokoro-82M
+!apt-get -qq -y install espeak-ng > /dev/null 2>&1
+!pip install -q phonemizer torch transformers scipy munch
+
+# 2️⃣ Build the model and load the default voicepack
+from models import build_model
+import torch
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+MODEL = build_model('kokoro-v0_19.pth', device)
+VOICE_NAME = [
+    'af', # Default voice is a 50-50 mix of Bella & Sarah
+    'af_bella', 'af_sarah', 'am_adam', 'am_michael',
+    'bf_emma', 'bf_isabella', 'bm_george', 'bm_lewis',
+    'af_nicole', 'af_sky',
+][0]
+VOICEPACK = torch.load(f'voices/{VOICE_NAME}.pt', weights_only=True).to(device)
+print(f'Loaded voice: {VOICE_NAME}')
+
+# 3️⃣ Call generate, which returns 24khz audio and the phonemes used
+from kokoro import generate
+text = "How could I know? It's an unanswerable question. Like asking an unborn child if they'll lead a good life. They haven't even been born."
+audio, out_ps = generate(MODEL, text, VOICEPACK, lang=VOICE_NAME[0])
+# Language is determined by the first letter of the VOICE_NAME:
+# 🇺🇸 'a' => American English => en-us
+# 🇬🇧 'b' => British English => en-gb
+
+# 4️⃣ Display the 24khz audio and print the output phonemes
+from IPython.display import display, Audio
+display(Audio(data=audio, rate=24000, autoplay=True))
+print(out_ps)
+```
+If you have trouble with `espeak-ng`, see this [github issue](https://github.com/bootphon/phonemizer/issues/44#issuecomment-1540885186). [Mac users also see this](https://huggingface.co/hexgrad/Kokoro-82M/discussions/12#677435d3d8ace1de46071489), and [Windows users see this](https://huggingface.co/hexgrad/Kokoro-82M/discussions/12#67742594fdeebf74f001ecfc).
+
+For ONNX usage, see [#14](https://huggingface.co/hexgrad/Kokoro-82M/discussions/14).
+
+### Model Facts
+
+No affiliation can be assumed between parties on different lines.
+
+**Architecture:**
+- StyleTTS 2: https://arxiv.org/abs/2306.07691
+- ISTFTNet: https://arxiv.org/abs/2203.02395
+- Decoder only: no diffusion, no encoder release
+
+**Architected by:** Li et al @ https://github.com/yl4579/StyleTTS2
+
+**Trained by**: `@rzvzn` on Discord
+
+**Supported Languages:** American English, British English
+
+**Model SHA256 Hash:** `3b0c392f87508da38fad3a2f9d94c359f1b657ebd2ef79f9d56d69503e470b0a`
+
+### Releases
+- 25 Dec 2024: Model v0.19, `af_bella`, `af_sarah`
+- 26 Dec 2024: `am_adam`, `am_michael`
+- 28 Dec 2024: `bf_emma`, `bf_isabella`, `bm_george`, `bm_lewis`
+- 30 Dec 2024: `af_nicole`
+- 31 Dec 2024: `af_sky`
+- 2 Jan 2025: ONNX v0.19 `ebef4245`
+
+### Licenses
+- Apache 2.0 weights in this repository
+- MIT inference code in [spaces/hexgrad/Kokoro-TTS](https://huggingface.co/spaces/hexgrad/Kokoro-TTS) adapted from [yl4579/StyleTTS2](https://github.com/yl4579/StyleTTS2)
+- GPLv3 dependency in [espeak-ng](https://github.com/espeak-ng/espeak-ng)
+
+The inference code was originally MIT licensed by the paper author. Note that this card applies only to this model, Kokoro. Original models published by the paper author can be found at [hf.co/yl4579](https://huggingface.co/yl4579).
+
+### Evaluation
+
+**Metric:** Elo rating
+
+**Leaderboard:** [hf.co/spaces/Pendrokar/TTS-Spaces-Arena](https://huggingface.co/spaces/Pendrokar/TTS-Spaces-Arena)
+
+![TTS-Spaces-Arena-25-Dec-2024](demo/TTS-Spaces-Arena-25-Dec-2024.png)
+
+The voice ranked in the Arena is a 50-50 mix of Bella and Sarah. For your convenience, this mix is included in this repository as `af.pt`, but you can trivially reproduce it like this:
+
+```py
+import torch
+bella = torch.load('voices/af_bella.pt', weights_only=True)
+sarah = torch.load('voices/af_sarah.pt', weights_only=True)
+af = torch.mean(torch.stack([bella, sarah]), dim=0)
+assert torch.equal(af, torch.load('voices/af.pt', weights_only=True))
+```
+
+### Training Details
+
+**Compute:** Kokoro was trained on A100 80GB vRAM instances rented from [Vast.ai](https://cloud.vast.ai/?ref_id=79907) (referral link). Vast was chosen over other compute providers due to its competitive on-demand hourly rates. The average hourly cost for the A100 80GB vRAM instances used for training was below $1/hr per GPU, which was around half the quoted rates from other providers at the time.
+
+**Data:** Kokoro was trained exclusively on **permissive/non-copyrighted audio data** and IPA phoneme labels. Examples of permissive/non-copyrighted audio include:
+- Public domain audio
+- Audio licensed under Apache, MIT, etc
+- Synthetic audio<sup>[1]</sup> generated by closed<sup>[2]</sup> TTS models from large providers<br/>
+[1] https://copyright.gov/ai/ai_policy_guidance.pdf<br/>
+[2] No synthetic audio from open TTS models or "custom voice clones"
+
+**Epochs:** Less than **20 epochs**
+
+**Total Dataset Size:** Less than **100 hours** of audio
+
+### Limitations
+
+Kokoro v0.19 is limited in some specific ways, due to its training set and/or architecture:
+- [Data] Lacks voice cloning capability, likely due to small <100h training set
+- [Arch] Relies on external g2p (espeak-ng), which introduces a class of g2p failure modes
+- [Data] Training dataset is mostly long-form reading and narration, not conversation
+- [Arch] At 82M params, Kokoro almost certainly falls to a well-trained 1B+ param diffusion transformer, or a many-billion-param MLLM like GPT-4o / Gemini 2.0 Flash
+- [Data] Multilingual capability is architecturally feasible, but training data is mostly English
+
+Refer to the [Philosophy discussion](https://huggingface.co/hexgrad/Kokoro-82M/discussions/5) to better understand these limitations.
+
+**Will the other voicepacks be released?** There is currently no release date scheduled for the other voicepacks, but in the meantime you can try them in the hosted demo at [hf.co/spaces/hexgrad/Kokoro-TTS](https://huggingface.co/spaces/hexgrad/Kokoro-TTS).
+
+### Acknowledgements
+- [@yl4579](https://huggingface.co/yl4579) for architecting StyleTTS 2
+- [@Pendrokar](https://huggingface.co/Pendrokar) for adding Kokoro as a contender in the TTS Spaces Arena
+
+### Model Card Contact
+
+`@rzvzn` on Discord. Server invite: https://discord.gg/QuGxSWBfQy
+
+<img src="https://static0.gamerantimages.com/wordpress/wp-content/uploads/2024/08/terminator-zero-41-1.jpg" width="400" alt="kokoro" />
+
+https://terminator.fandom.com/wiki/Kokoro

config.json

@@ -0,0 +1,26 @@
+{
+  "decoder": {
+    "type": "istftnet",
+    "upsample_kernel_sizes": [20, 12],
+    "upsample_rates": [10, 6],
+    "gen_istft_hop_size": 5,
+    "gen_istft_n_fft": 20,
+    "resblock_dilation_sizes": [
+      [1, 3, 5],
+      [1, 3, 5],
+      [1, 3, 5]
+    ],
+    "resblock_kernel_sizes": [3, 7, 11],
+    "upsample_initial_channel": 512
+  },
+  "dim_in": 64,
+  "dropout": 0.2,
+  "hidden_dim": 512,
+  "max_conv_dim": 512,
+  "max_dur": 50,
+  "multispeaker": true,
+  "n_layer": 3,
+  "n_mels": 80,
+  "n_token": 178,
+  "style_dim": 128
+}

demo/HEARME.txt

@@ -0,0 +1,47 @@
+Kokoro is a frontier TTS model for its size of 82 million parameters.
+
+On the 25th of December, 2024, Kokoro v0 point 19 weights were permissively released in full fp32 precision along with 2 voicepacks (Bella and Sarah), all under an Apache 2 license.
+
+At the time of release, Kokoro v0 point 19 was the number 1 ranked model in TTS Spaces Arena. With 82 million parameters trained for under 20 epics on under 100 total hours of audio, Kokoro achieved higher Eelo in this single-voice Arena setting, over larger models. Kokoro's ability to top this Eelo ladder using relatively low compute and data, suggests that the scaling law for traditional TTS models might have a steeper slope than previously expected.
+
+Licenses. Apache 2 weights in this repository. MIT inference code. GPLv3 dependency in espeak NG.
+
+The inference code was originally MIT licensed by the paper author. Note that this card applies only to this model, Kokoro.
+
+Evaluation. Metric: Eelo rating. Leaderboard: TTS Spaces Arena.
+
+The voice ranked in the Arena is a 50 50 mix of Bella and Sarah. For your convenience, this mix is included in this repository as A-F dot PT, but you can trivially re-produce it.
+
+Training Details.
+
+Compute: Kokoro was trained on "A100 80GB v-ram instances" rented from Vast.ai. Vast was chosen over other compute providers due to its competitive on-demand hourly rates. The average hourly cost for the A100 80GB v-ram instances used for training was below $1 per hour per GPU, which was around half the quoted rates from other providers at the time.
+
+Data: Kokoro was trained exclusively on permissive non-copyrighted audio data and IPA phoneme labels. Examples of permissive non-copyrighted audio include:
+
+Public domain audio. Audio licensed under Apache, MIT, etc.
+
+Synthetic audio[1] generated by closed[2] TTS models from large providers.
+
+Epics: Less than 20 Epics. Total Dataset Size: Less than 100 hours of audio.
+
+Limitations. Kokoro v0 point 19 is limited in some ways, in its training set and architecture:
+
+Lacks voice cloning capability, likely due to small, under 100 hour training set.
+
+Relies on external g2p, which introduces a class of g2p failure modes.
+
+Training dataset is mostly long-form reading and narration, not conversation.
+
+At 82 million parameters, Kokoro almost certainly falls to a well-trained 1B+ parameter diffusion transformer, or a many-billion-parameter M LLM like GPT 4o or Gemini 2 Flash.
+
+Multilingual capability is architecturally feasible, but training data is almost entirely English.
+
+Will the other voicepacks be released?
+
+There is currently no release date scheduled for the other voicepacks, but in the meantime you can try them in the hosted demo.
+
+Acknowledgements. yL4 5 7 9 for architecting StyleTTS 2.
+
+Pendrokar for adding Kokoro as a contender in the TTS Spaces Arena.
+
+Model Card Contact. @rzvzn on Discord.

demo/HEARME.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98b884082db74c250b3cecda78341d1724c66727c0391b29a0160af918eccdb3
+size 11198508

demo/af_sky.txt

@@ -0,0 +1,11 @@
+Last September, I received an offer from Sam Altman, who wanted to hire me to voice the current ChatGPT 4 system. He told me that he felt that by my voicing the system, I could bridge the gap between tech companies and creatives and help consumers to feel comfortable with the seismic shift concerning humans and AI. He said he felt that my voice would be comforting to people.
+
+After much consideration and for personal reasons, I declined the offer. Nine months later, my friends, family and the general public all noted how much the newest system named Sky sounded like me.
+
+When I heard the released demo, I was shocked, angered and in disbelief that Mr. Altman would pursue a voice that sounded so eerily similar to mine that my closest friends and news ou'tlits could not tell the difference. Mr. Altman even insinuated that the similarity was intentional, tweeting a single word — hur — a reference to the film in which I voiced a chat system, Samantha, who forms an intimate relationship with a human.
+
+Two days before the ChatGPT 4 demo was released, Mr. Altman contacted my agent, asking me to reconsider. Before we could connect, the system was out there.
+
+As a result of their actions, I was forced to hire legal counsel, who wrote two letters to Mr. Altman and OpenAI, setting out what they had done and asking them to detail the exact process by which they created the Sky voice. Consequently, OpenAI reluctantly agreed to take down the Sky voice.
+
+In a time when we are all grappling with deepfakes and the protection of our own likeness, our own work, our own identities, I believe these are questions that deserve absolute clarity. I look forward to resolution in the form of transparency and the passage of appropriate legislation to help ensure that individual rights are protected.

demo/af_sky.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce36292bf868aa5f15931f3d81a9f46cc35ea76372e618a5e4453c9542e5ad7e
+size 5486636

demo/restoring-sky.md

@@ -0,0 +1,42 @@
+# Restoring Sky & reflecting on Kokoro
+
+<img src="https://static0.gamerantimages.com/wordpress/wp-content/uploads/2024/08/terminator-zero-41-1.jpg" width="400" alt="kokoro" />
+
+For those who don't know, [Kokoro](https://huggingface.co/hexgrad/Kokoro-82M) is an Apache TTS model that uses a skinny version of the open [StyleTTS 2](https://github.com/yl4579/StyleTTS2/tree/main) architecture.
+
+Based on leaderboard [Elo rating](https://huggingface.co/hexgrad/Kokoro-82M#evaluation) (prior to getting [review bombed](https://huggingface.co/datasets/Pendrokar/TTS_Arena/discussions/2)), Kokoro appears to do more with less, a theme that is surely [top-of-mind](https://huggingface.co/deepseek-ai/DeepSeek-V3) for many. It's peak performance on specific voices is comparable or better than much larger models, but it has not yet been trained on enough data to effectively zero-shot out of distribution (aka voice cloning).
+
+Tonight on NYE, `af_sky` joins Kokoro's roster of downloadable voices. This follows last night's quiet release of `af_nicole`, and an additional 8 voices are currently available: 2F 2M voices each for American & British English.
+
+Nicole in particular was trained on ~10 hours of synthetic data, and demonstrates that you _can_ include unique speaking styles in a general-purpose TTS model without affecting the stock voices (even in a low data small model): a good sign for scalability.
+
+Sky is interesting because it is the voice that ScarJo [got OpenAI to take down](https://x.com/OpenAI/status/1792443575839678909), so new training data cannot be generated. However, OpenAI did not remove 2023 samples of Sky from their [blog post](https://openai.com/index/chatgpt-can-now-see-hear-and-speak/), and along with a few seconds lying around various other parts of the internet, we can cobble together about 3 minutes of 2023 Sky.
+
+```sh
+wget https://cdn.openai.com/new-voice-and-image-capabilities-in-chatgpt/hd/story-sky.mp3
+wget https://cdn.openai.com/new-voice-and-image-capabilities-in-chatgpt/hd/recipe-sky.mp3
+wget https://cdn.openai.com/new-voice-and-image-capabilities-in-chatgpt/hd/speech-sky.mp3
+wget https://cdn.openai.com/new-voice-and-image-capabilities-in-chatgpt/hd/poem-sky.mp3
+wget https://cdn.openai.com/new-voice-and-image-capabilities-in-chatgpt/hd/info-sky.mp3
+```
+
+To be clear, this is not the first attempt to reconstruct Sky. On X, Benjamin De Kraker posted:
+> Here's the official statement released by Scarlett Johansson, detailing OpenAI's alleged illegal usage of her voice...
+> ...read by the Sky AI voice, because irony.
+> https://x.com/BenjaminDEKR/status/1792693868497871086
+
+and in the replies, he [stated](https://x.com/BenjaminDEKR/status/1792714347275501595):
+> It's an ElevenLabs clone I made based on Sky audio before they removed it. Not perfect.
+
+Here is `Kokoro/af_sky`'s rendition of the same:
+<audio controls><source src="https://huggingface.co/hexgrad/Kokoro-82M/resolve/main/demo/af_sky.wav" type="audio/wav"></audio>
+
+A crude reconstruction, but the model that produced that voice is Apache FOSS that can be downloaded from HF and run locally. You can reproduce the above by dragging the [text script](https://huggingface.co/hexgrad/Kokoro-82M/blob/main/demo/af_sky.txt) (note a handful of modified chars for better delivery) into the "Long Form" tab of this [hosted demo](https://huggingface.co/spaces/hexgrad/Kokoro-TTS), or you can download the [model weights](https://huggingface.co/hexgrad/Kokoro-82M), install dependencies and DIY.
+
+Sky shows that it is possible to reconstruct a voice—maybe a shadow of its former self, but a reconstruction nonetheless—from fairly little training data.
+
+### What's next
+
+Kokoro is a good start, but I can think of some tricks that might make it better, beginning with better data. More on this in another article.
+
+Feel free to check out [Kokoro's weights](https://huggingface.co/hexgrad/Kokoro-82M), try out a no-install [hosted demo](https://huggingface.co/spaces/hexgrad/Kokoro-TTS), and/or [join the Discord](https://discord.gg/QuGxSWBfQy).

fp16/halve.py

@@ -0,0 +1,17 @@
+from hashlib import sha256
+from pathlib import Path
+import torch
+
+path = Path(__file__).parent.parent / 'kokoro-v0_19.pth'
+assert path.exists(), f'No model pth found at {path}'
+
+net = torch.load(path, map_location='cpu', weights_only=True)['net']
+for a in net:
+    for b in net[a]:
+        net[a][b] = net[a][b].half()
+
+torch.save(dict(net=net), 'kokoro-v0_19-half.pth')
+with open('kokoro-v0_19-half.pth', 'rb') as rb:
+    h = sha256(rb.read()).hexdigest()
+
+assert h == '70cbf37f84610967f2ca72dadb95456fdd8b6c72cdd6dc7372c50f525889ff0c', h

fp16/kokoro-v0_19-half.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70cbf37f84610967f2ca72dadb95456fdd8b6c72cdd6dc7372c50f525889ff0c
+size 163731194

istftnet.py

@@ -0,0 +1,523 @@
+# https://github.com/yl4579/StyleTTS2/blob/main/Modules/istftnet.py
+from scipy.signal import get_window
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils import weight_norm, remove_weight_norm
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# https://github.com/yl4579/StyleTTS2/blob/main/Modules/utils.py
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size*dilation - dilation)/2)
+
+LRELU_SLOPE = 0.1
+
+class AdaIN1d(nn.Module):
+    def __init__(self, style_dim, num_features):
+        super().__init__()
+        self.norm = nn.InstanceNorm1d(num_features, affine=False)
+        self.fc = nn.Linear(style_dim, num_features*2)
+
+    def forward(self, x, s):
+        h = self.fc(s)
+        h = h.view(h.size(0), h.size(1), 1)
+        gamma, beta = torch.chunk(h, chunks=2, dim=1)
+        return (1 + gamma) * self.norm(x) + beta
+
+class AdaINResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5), style_dim=64):
+        super(AdaINResBlock1, self).__init__()
+        self.convs1 = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
+                               padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
+                               padding=get_padding(kernel_size, dilation[1]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
+                               padding=get_padding(kernel_size, dilation[2])))
+        ])
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1)))
+        ])
+        self.convs2.apply(init_weights)
+        
+        self.adain1 = nn.ModuleList([
+            AdaIN1d(style_dim, channels),
+            AdaIN1d(style_dim, channels),
+            AdaIN1d(style_dim, channels),
+        ])
+        
+        self.adain2 = nn.ModuleList([
+            AdaIN1d(style_dim, channels),
+            AdaIN1d(style_dim, channels),
+            AdaIN1d(style_dim, channels),
+        ])
+        
+        self.alpha1 = nn.ParameterList([nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs1))])
+        self.alpha2 = nn.ParameterList([nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs2))])
+
+
+    def forward(self, x, s):
+        for c1, c2, n1, n2, a1, a2 in zip(self.convs1, self.convs2, self.adain1, self.adain2, self.alpha1, self.alpha2):
+            xt = n1(x, s)
+            xt = xt + (1 / a1) * (torch.sin(a1 * xt) ** 2)  # Snake1D
+            xt = c1(xt)
+            xt = n2(xt, s)
+            xt = xt + (1 / a2) * (torch.sin(a2 * xt) ** 2)  # Snake1D
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+            
+class TorchSTFT(torch.nn.Module):
+    def __init__(self, filter_length=800, hop_length=200, win_length=800, window='hann'):
+        super().__init__()
+        self.filter_length = filter_length
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.window = torch.from_numpy(get_window(window, win_length, fftbins=True).astype(np.float32))
+
+    def transform(self, input_data):
+        forward_transform = torch.stft(
+            input_data,
+            self.filter_length, self.hop_length, self.win_length, window=self.window.to(input_data.device),
+            return_complex=True)
+
+        return torch.abs(forward_transform), torch.angle(forward_transform)
+
+    def inverse(self, magnitude, phase):
+        inverse_transform = torch.istft(
+            magnitude * torch.exp(phase * 1j),
+            self.filter_length, self.hop_length, self.win_length, window=self.window.to(magnitude.device))
+
+        return inverse_transform.unsqueeze(-2)  # unsqueeze to stay consistent with conv_transpose1d implementation
+
+    def forward(self, input_data):
+        self.magnitude, self.phase = self.transform(input_data)
+        reconstruction = self.inverse(self.magnitude, self.phase)
+        return reconstruction
+    
+class SineGen(torch.nn.Module):
+    """ Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(self, samp_rate, upsample_scale, harmonic_num=0,
+                 sine_amp=0.1, noise_std=0.003,
+                 voiced_threshold=0,
+                 flag_for_pulse=False):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.dim = self.harmonic_num + 1
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+        self.flag_for_pulse = flag_for_pulse
+        self.upsample_scale = upsample_scale
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = (f0 > self.voiced_threshold).type(torch.float32)
+        return uv
+
+    def _f02sine(self, f0_values):
+        """ f0_values: (batchsize, length, dim)
+            where dim indicates fundamental tone and overtones
+        """
+        # convert to F0 in rad. The interger part n can be ignored
+        # because 2 * np.pi * n doesn't affect phase
+        rad_values = (f0_values / self.sampling_rate) % 1
+
+        # initial phase noise (no noise for fundamental component)
+        rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], \
+                              device=f0_values.device)
+        rand_ini[:, 0] = 0
+        rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+
+        # instantanouse phase sine[t] = sin(2*pi \sum_i=1 ^{t} rad)
+        if not self.flag_for_pulse:
+#             # for normal case
+
+#             # To prevent torch.cumsum numerical overflow,
+#             # it is necessary to add -1 whenever \sum_k=1^n rad_value_k > 1.
+#             # Buffer tmp_over_one_idx indicates the time step to add -1.
+#             # This will not change F0 of sine because (x-1) * 2*pi = x * 2*pi
+#             tmp_over_one = torch.cumsum(rad_values, 1) % 1
+#             tmp_over_one_idx = (padDiff(tmp_over_one)) < 0
+#             cumsum_shift = torch.zeros_like(rad_values)
+#             cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+
+#             phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
+            rad_values = torch.nn.functional.interpolate(rad_values.transpose(1, 2), 
+                                                         scale_factor=1/self.upsample_scale, 
+                                                         mode="linear").transpose(1, 2)
+    
+#             tmp_over_one = torch.cumsum(rad_values, 1) % 1
+#             tmp_over_one_idx = (padDiff(tmp_over_one)) < 0
+#             cumsum_shift = torch.zeros_like(rad_values)
+#             cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+    
+            phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
+            phase = torch.nn.functional.interpolate(phase.transpose(1, 2) * self.upsample_scale, 
+                                                    scale_factor=self.upsample_scale, mode="linear").transpose(1, 2)
+            sines = torch.sin(phase)
+            
+        else:
+            # If necessary, make sure that the first time step of every
+            # voiced segments is sin(pi) or cos(0)
+            # This is used for pulse-train generation
+
+            # identify the last time step in unvoiced segments
+            uv = self._f02uv(f0_values)
+            uv_1 = torch.roll(uv, shifts=-1, dims=1)
+            uv_1[:, -1, :] = 1
+            u_loc = (uv < 1) * (uv_1 > 0)
+
+            # get the instantanouse phase
+            tmp_cumsum = torch.cumsum(rad_values, dim=1)
+            # different batch needs to be processed differently
+            for idx in range(f0_values.shape[0]):
+                temp_sum = tmp_cumsum[idx, u_loc[idx, :, 0], :]
+                temp_sum[1:, :] = temp_sum[1:, :] - temp_sum[0:-1, :]
+                # stores the accumulation of i.phase within
+                # each voiced segments
+                tmp_cumsum[idx, :, :] = 0
+                tmp_cumsum[idx, u_loc[idx, :, 0], :] = temp_sum
+
+            # rad_values - tmp_cumsum: remove the accumulation of i.phase
+            # within the previous voiced segment.
+            i_phase = torch.cumsum(rad_values - tmp_cumsum, dim=1)
+
+            # get the sines
+            sines = torch.cos(i_phase * 2 * np.pi)
+        return sines
+
+    def forward(self, f0):
+        """ sine_tensor, uv = forward(f0)
+        input F0: tensor(batchsize=1, length, dim=1)
+                  f0 for unvoiced steps should be 0
+        output sine_tensor: tensor(batchsize=1, length, dim)
+        output uv: tensor(batchsize=1, length, 1)
+        """
+        f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim,
+                             device=f0.device)
+        # fundamental component
+        fn = torch.multiply(f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device))
+
+        # generate sine waveforms
+        sine_waves = self._f02sine(fn) * self.sine_amp
+
+        # generate uv signal
+        # uv = torch.ones(f0.shape)
+        # uv = uv * (f0 > self.voiced_threshold)
+        uv = self._f02uv(f0)
+
+        # noise: for unvoiced should be similar to sine_amp
+        #        std = self.sine_amp/3 -> max value ~ self.sine_amp
+        # .       for voiced regions is self.noise_std
+        noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+        noise = noise_amp * torch.randn_like(sine_waves)
+
+        # first: set the unvoiced part to 0 by uv
+        # then: additive noise
+        sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """ SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(sampling_rate, upsample_scale, harmonic_num,
+                                 sine_amp, add_noise_std, voiced_threshod)
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x):
+        """
+        Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+        F0_sampled (batchsize, length, 1)
+        Sine_source (batchsize, length, 1)
+        noise_source (batchsize, length 1)
+        """
+        # source for harmonic branch
+        with torch.no_grad():
+            sine_wavs, uv, _ = self.l_sin_gen(x)
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+
+        # source for noise branch, in the same shape as uv
+        noise = torch.randn_like(uv) * self.sine_amp / 3
+        return sine_merge, noise, uv
+def padDiff(x):
+    return F.pad(F.pad(x, (0,0,-1,1), 'constant', 0) - x, (0,0,0,-1), 'constant', 0)
+
+    
+class Generator(torch.nn.Module):
+    def __init__(self, style_dim, resblock_kernel_sizes, upsample_rates, upsample_initial_channel, resblock_dilation_sizes, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size):
+        super(Generator, self).__init__()
+
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        resblock = AdaINResBlock1
+
+        self.m_source = SourceModuleHnNSF(
+                    sampling_rate=24000,
+                    upsample_scale=np.prod(upsample_rates) * gen_istft_hop_size,
+                    harmonic_num=8, voiced_threshod=10)
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates) * gen_istft_hop_size)
+        self.noise_convs = nn.ModuleList()
+        self.noise_res = nn.ModuleList()
+        
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(weight_norm(
+                ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
+                                k, u, padding=(k-u)//2)))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel//(2**(i+1))
+            for j, (k, d) in enumerate(zip(resblock_kernel_sizes,resblock_dilation_sizes)):
+                self.resblocks.append(resblock(ch, k, d, style_dim))
+                
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            
+            if i + 1 < len(upsample_rates):  #
+                stride_f0 = np.prod(upsample_rates[i + 1:])
+                self.noise_convs.append(Conv1d(
+                    gen_istft_n_fft + 2, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=(stride_f0+1) // 2))
+                self.noise_res.append(resblock(c_cur, 7, [1,3,5], style_dim))
+            else:
+                self.noise_convs.append(Conv1d(gen_istft_n_fft + 2, c_cur, kernel_size=1))
+                self.noise_res.append(resblock(c_cur, 11, [1,3,5], style_dim))
+                
+                
+        self.post_n_fft = gen_istft_n_fft
+        self.conv_post = weight_norm(Conv1d(ch, self.post_n_fft + 2, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+        self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
+        self.stft = TorchSTFT(filter_length=gen_istft_n_fft, hop_length=gen_istft_hop_size, win_length=gen_istft_n_fft)
+        
+        
+    def forward(self, x, s, f0):
+        with torch.no_grad():
+            f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
+
+            har_source, noi_source, uv = self.m_source(f0)
+            har_source = har_source.transpose(1, 2).squeeze(1)
+            har_spec, har_phase = self.stft.transform(har_source)
+            har = torch.cat([har_spec, har_phase], dim=1)
+        
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x_source = self.noise_convs[i](har)
+            x_source = self.noise_res[i](x_source, s)
+
+            x = self.ups[i](x)
+            if i == self.num_upsamples - 1:
+                x = self.reflection_pad(x)
+
+            x = x + x_source
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i*self.num_kernels+j](x, s)
+                else:
+                    xs += self.resblocks[i*self.num_kernels+j](x, s)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        spec = torch.exp(x[:,:self.post_n_fft // 2 + 1, :])
+        phase = torch.sin(x[:, self.post_n_fft // 2 + 1:, :])
+        return self.stft.inverse(spec, phase)
+    
+    def fw_phase(self, x, s):
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i*self.num_kernels+j](x, s)
+                else:
+                    xs += self.resblocks[i*self.num_kernels+j](x, s)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.reflection_pad(x)
+        x = self.conv_post(x)
+        spec = torch.exp(x[:,:self.post_n_fft // 2 + 1, :])
+        phase = torch.sin(x[:, self.post_n_fft // 2 + 1:, :])
+        return spec, phase
+
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+        
+class AdainResBlk1d(nn.Module):
+    def __init__(self, dim_in, dim_out, style_dim=64, actv=nn.LeakyReLU(0.2),
+                 upsample='none', dropout_p=0.0):
+        super().__init__()
+        self.actv = actv
+        self.upsample_type = upsample
+        self.upsample = UpSample1d(upsample)
+        self.learned_sc = dim_in != dim_out
+        self._build_weights(dim_in, dim_out, style_dim)
+        self.dropout = nn.Dropout(dropout_p)
+        
+        if upsample == 'none':
+            self.pool = nn.Identity()
+        else:
+            self.pool = weight_norm(nn.ConvTranspose1d(dim_in, dim_in, kernel_size=3, stride=2, groups=dim_in, padding=1, output_padding=1))
+        
+        
+    def _build_weights(self, dim_in, dim_out, style_dim):
+        self.conv1 = weight_norm(nn.Conv1d(dim_in, dim_out, 3, 1, 1))
+        self.conv2 = weight_norm(nn.Conv1d(dim_out, dim_out, 3, 1, 1))
+        self.norm1 = AdaIN1d(style_dim, dim_in)
+        self.norm2 = AdaIN1d(style_dim, dim_out)
+        if self.learned_sc:
+            self.conv1x1 = weight_norm(nn.Conv1d(dim_in, dim_out, 1, 1, 0, bias=False))
+
+    def _shortcut(self, x):
+        x = self.upsample(x)
+        if self.learned_sc:
+            x = self.conv1x1(x)
+        return x
+
+    def _residual(self, x, s):
+        x = self.norm1(x, s)
+        x = self.actv(x)
+        x = self.pool(x)
+        x = self.conv1(self.dropout(x))
+        x = self.norm2(x, s)
+        x = self.actv(x)
+        x = self.conv2(self.dropout(x))
+        return x
+
+    def forward(self, x, s):
+        out = self._residual(x, s)
+        out = (out + self._shortcut(x)) / np.sqrt(2)
+        return out
+    
+class UpSample1d(nn.Module):
+    def __init__(self, layer_type):
+        super().__init__()
+        self.layer_type = layer_type
+
+    def forward(self, x):
+        if self.layer_type == 'none':
+            return x
+        else:
+            return F.interpolate(x, scale_factor=2, mode='nearest')
+
+class Decoder(nn.Module):
+    def __init__(self, dim_in=512, F0_channel=512, style_dim=64, dim_out=80, 
+                resblock_kernel_sizes = [3,7,11],
+                upsample_rates = [10, 6],
+                upsample_initial_channel=512,
+                resblock_dilation_sizes=[[1,3,5], [1,3,5], [1,3,5]],
+                upsample_kernel_sizes=[20, 12], 
+                gen_istft_n_fft=20, gen_istft_hop_size=5):
+        super().__init__()
+        
+        self.decode = nn.ModuleList()
+        
+        self.encode = AdainResBlk1d(dim_in + 2, 1024, style_dim)
+        
+        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
+        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
+        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
+        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 512, style_dim, upsample=True))
+
+        self.F0_conv = weight_norm(nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1))
+        
+        self.N_conv = weight_norm(nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1))
+        
+        self.asr_res = nn.Sequential(
+            weight_norm(nn.Conv1d(512, 64, kernel_size=1)),
+        )
+        
+        
+        self.generator = Generator(style_dim, resblock_kernel_sizes, upsample_rates, 
+                                   upsample_initial_channel, resblock_dilation_sizes, 
+                                   upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size)
+        
+    def forward(self, asr, F0_curve, N, s):
+        F0 = self.F0_conv(F0_curve.unsqueeze(1))
+        N = self.N_conv(N.unsqueeze(1))
+        
+        x = torch.cat([asr, F0, N], axis=1)
+        x = self.encode(x, s)
+        
+        asr_res = self.asr_res(asr)
+        
+        res = True
+        for block in self.decode:
+            if res:
+                x = torch.cat([x, asr_res, F0, N], axis=1)
+            x = block(x, s)
+            if block.upsample_type != "none":
+                res = False
+                
+        x = self.generator(x, s, F0_curve)
+        return x

kokoro-v0_19.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebef42457f7efee9b60b4f1d5aec7692f2925923948a0d7a2a49d2c9edf57e49
+size 345554732

kokoro-v0_19.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b0c392f87508da38fad3a2f9d94c359f1b657ebd2ef79f9d56d69503e470b0a
+size 327211206

kokoro.py

@@ -0,0 +1,149 @@
+import phonemizer
+import re
+import torch
+
+def split_num(num):
+    num = num.group()
+    if '.' in num:
+        return num
+    elif ':' in num:
+        h, m = [int(n) for n in num.split(':')]
+        if m == 0:
+            return f"{h} o'clock"
+        elif m < 10:
+            return f'{h} oh {m}'
+        return f'{h} {m}'
+    year = int(num[:4])
+    if year < 1100 or year % 1000 < 10:
+        return num
+    left, right = num[:2], int(num[2:4])
+    s = 's' if num.endswith('s') else ''
+    if 100 <= year % 1000 <= 999:
+        if right == 0:
+            return f'{left} hundred{s}'
+        elif right < 10:
+            return f'{left} oh {right}{s}'
+    return f'{left} {right}{s}'
+
+def flip_money(m):
+    m = m.group()
+    bill = 'dollar' if m[0] == '$' else 'pound'
+    if m[-1].isalpha():
+        return f'{m[1:]} {bill}s'
+    elif '.' not in m:
+        s = '' if m[1:] == '1' else 's'
+        return f'{m[1:]} {bill}{s}'
+    b, c = m[1:].split('.')
+    s = '' if b == '1' else 's'
+    c = int(c.ljust(2, '0'))
+    coins = f"cent{'' if c == 1 else 's'}" if m[0] == '$' else ('penny' if c == 1 else 'pence')
+    return f'{b} {bill}{s} and {c} {coins}'
+
+def point_num(num):
+    a, b = num.group().split('.')
+    return ' point '.join([a, ' '.join(b)])
+
+def normalize_text(text):
+    text = text.replace(chr(8216), "'").replace(chr(8217), "'")
+    text = text.replace('«', chr(8220)).replace('»', chr(8221))
+    text = text.replace(chr(8220), '"').replace(chr(8221), '"')
+    text = text.replace('(', '«').replace(')', '»')
+    for a, b in zip('、。！，：；？', ',.!,:;?'):
+        text = text.replace(a, b+' ')
+    text = re.sub(r'[^\S \n]', ' ', text)
+    text = re.sub(r'  +', ' ', text)
+    text = re.sub(r'(?<=\n) +(?=\n)', '', text)
+    text = re.sub(r'\bD[Rr]\.(?= [A-Z])', 'Doctor', text)
+    text = re.sub(r'\b(?:Mr\.|MR\.(?= [A-Z]))', 'Mister', text)
+    text = re.sub(r'\b(?:Ms\.|MS\.(?= [A-Z]))', 'Miss', text)
+    text = re.sub(r'\b(?:Mrs\.|MRS\.(?= [A-Z]))', 'Mrs', text)
+    text = re.sub(r'\betc\.(?! [A-Z])', 'etc', text)
+    text = re.sub(r'(?i)\b(y)eah?\b', r"\1e'a", text)
+    text = re.sub(r'\d*\.\d+|\b\d{4}s?\b|(?<!:)\b(?:[1-9]|1[0-2]):[0-5]\d\b(?!:)', split_num, text)
+    text = re.sub(r'(?<=\d),(?=\d)', '', text)
+    text = re.sub(r'(?i)[$£]\d+(?:\.\d+)?(?: hundred| thousand| (?:[bm]|tr)illion)*\b|[$£]\d+\.\d\d?\b', flip_money, text)
+    text = re.sub(r'\d*\.\d+', point_num, text)
+    text = re.sub(r'(?<=\d)-(?=\d)', ' to ', text)
+    text = re.sub(r'(?<=\d)S', ' S', text)
+    text = re.sub(r"(?<=[BCDFGHJ-NP-TV-Z])'?s\b", "'S", text)
+    text = re.sub(r"(?<=X')S\b", 's', text)
+    text = re.sub(r'(?:[A-Za-z]\.){2,} [a-z]', lambda m: m.group().replace('.', '-'), text)
+    text = re.sub(r'(?i)(?<=[A-Z])\.(?=[A-Z])', '-', text)
+    return text.strip()
+
+def get_vocab():
+    _pad = "$"
+    _punctuation = ';:,.!?¡¿—…"«»“” '
+    _letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
+    _letters_ipa = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
+    symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa)
+    dicts = {}
+    for i in range(len((symbols))):
+        dicts[symbols[i]] = i
+    return dicts
+
+VOCAB = get_vocab()
+def tokenize(ps):
+    return [i for i in map(VOCAB.get, ps) if i is not None]
+
+phonemizers = dict(
+    a=phonemizer.backend.EspeakBackend(language='en-us', preserve_punctuation=True, with_stress=True),
+    b=phonemizer.backend.EspeakBackend(language='en-gb', preserve_punctuation=True, with_stress=True),
+)
+def phonemize(text, lang, norm=True):
+    if norm:
+        text = normalize_text(text)
+    ps = phonemizers[lang].phonemize([text])
+    ps = ps[0] if ps else ''
+    # https://en.wiktionary.org/wiki/kokoro#English
+    ps = ps.replace('kəkˈoːɹoʊ', 'kˈoʊkəɹoʊ').replace('kəkˈɔːɹəʊ', 'kˈəʊkəɹəʊ')
+    ps = ps.replace('ʲ', 'j').replace('r', 'ɹ').replace('x', 'k').replace('ɬ', 'l')
+    ps = re.sub(r'(?<=[a-zɹː])(?=hˈʌndɹɪd)', ' ', ps)
+    ps = re.sub(r' z(?=[;:,.!?¡¿—…"«»“” ]|$)', 'z', ps)
+    if lang == 'a':
+        ps = re.sub(r'(?<=nˈaɪn)ti(?!ː)', 'di', ps)
+    ps = ''.join(filter(lambda p: p in VOCAB, ps))
+    return ps.strip()
+
+def length_to_mask(lengths):
+    mask = torch.arange(lengths.max()).unsqueeze(0).expand(lengths.shape[0], -1).type_as(lengths)
+    mask = torch.gt(mask+1, lengths.unsqueeze(1))
+    return mask
+
+@torch.no_grad()
+def forward(model, tokens, ref_s, speed):
+    device = ref_s.device
+    tokens = torch.LongTensor([[0, *tokens, 0]]).to(device)
+    input_lengths = torch.LongTensor([tokens.shape[-1]]).to(device)
+    text_mask = length_to_mask(input_lengths).to(device)
+    bert_dur = model.bert(tokens, attention_mask=(~text_mask).int())
+    d_en = model.bert_encoder(bert_dur).transpose(-1, -2)
+    s = ref_s[:, 128:]
+    d = model.predictor.text_encoder(d_en, s, input_lengths, text_mask)
+    x, _ = model.predictor.lstm(d)
+    duration = model.predictor.duration_proj(x)
+    duration = torch.sigmoid(duration).sum(axis=-1) / speed
+    pred_dur = torch.round(duration).clamp(min=1).long()
+    pred_aln_trg = torch.zeros(input_lengths, pred_dur.sum().item())
+    c_frame = 0
+    for i in range(pred_aln_trg.size(0)):
+        pred_aln_trg[i, c_frame:c_frame + pred_dur[0,i].item()] = 1
+        c_frame += pred_dur[0,i].item()
+    en = d.transpose(-1, -2) @ pred_aln_trg.unsqueeze(0).to(device)
+    F0_pred, N_pred = model.predictor.F0Ntrain(en, s)
+    t_en = model.text_encoder(tokens, input_lengths, text_mask)
+    asr = t_en @ pred_aln_trg.unsqueeze(0).to(device)
+    return model.decoder(asr, F0_pred, N_pred, ref_s[:, :128]).squeeze().cpu().numpy()
+
+def generate(model, text, voicepack, lang='a', speed=1, ps=None):
+    ps = ps or phonemize(text, lang)
+    tokens = tokenize(ps)
+    if not tokens:
+        return None
+    elif len(tokens) > 510:
+        tokens = tokens[:510]
+        print('Truncated to 510 tokens')
+    ref_s = voicepack[len(tokens)]
+    out = forward(model, tokens, ref_s, speed)
+    ps = ''.join(next(k for k, v in VOCAB.items() if i == v) for i in tokens)
+    return out, ps

models.py

@@ -0,0 +1,372 @@
+# https://github.com/yl4579/StyleTTS2/blob/main/models.py
+from istftnet import AdaIN1d, Decoder
+from munch import Munch
+from pathlib import Path
+from plbert import load_plbert
+from torch.nn.utils import weight_norm, spectral_norm
+import json
+import numpy as np
+import os
+import os.path as osp
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class LinearNorm(torch.nn.Module):
+    def __init__(self, in_dim, out_dim, bias=True, w_init_gain='linear'):
+        super(LinearNorm, self).__init__()
+        self.linear_layer = torch.nn.Linear(in_dim, out_dim, bias=bias)
+
+        torch.nn.init.xavier_uniform_(
+            self.linear_layer.weight,
+            gain=torch.nn.init.calculate_gain(w_init_gain))
+
+    def forward(self, x):
+        return self.linear_layer(x)
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+    
+class TextEncoder(nn.Module):
+    def __init__(self, channels, kernel_size, depth, n_symbols, actv=nn.LeakyReLU(0.2)):
+        super().__init__()
+        self.embedding = nn.Embedding(n_symbols, channels)
+
+        padding = (kernel_size - 1) // 2
+        self.cnn = nn.ModuleList()
+        for _ in range(depth):
+            self.cnn.append(nn.Sequential(
+                weight_norm(nn.Conv1d(channels, channels, kernel_size=kernel_size, padding=padding)),
+                LayerNorm(channels),
+                actv,
+                nn.Dropout(0.2),
+            ))
+        # self.cnn = nn.Sequential(*self.cnn)
+
+        self.lstm = nn.LSTM(channels, channels//2, 1, batch_first=True, bidirectional=True)
+
+    def forward(self, x, input_lengths, m):
+        x = self.embedding(x)  # [B, T, emb]
+        x = x.transpose(1, 2)  # [B, emb, T]
+        m = m.to(input_lengths.device).unsqueeze(1)
+        x.masked_fill_(m, 0.0)
+        
+        for c in self.cnn:
+            x = c(x)
+            x.masked_fill_(m, 0.0)
+            
+        x = x.transpose(1, 2)  # [B, T, chn]
+
+        input_lengths = input_lengths.cpu().numpy()
+        x = nn.utils.rnn.pack_padded_sequence(
+            x, input_lengths, batch_first=True, enforce_sorted=False)
+
+        self.lstm.flatten_parameters()
+        x, _ = self.lstm(x)
+        x, _ = nn.utils.rnn.pad_packed_sequence(
+            x, batch_first=True)
+                
+        x = x.transpose(-1, -2)
+        x_pad = torch.zeros([x.shape[0], x.shape[1], m.shape[-1]])
+
+        x_pad[:, :, :x.shape[-1]] = x
+        x = x_pad.to(x.device)
+        
+        x.masked_fill_(m, 0.0)
+        
+        return x
+
+    def inference(self, x):
+        x = self.embedding(x)
+        x = x.transpose(1, 2)
+        x = self.cnn(x)
+        x = x.transpose(1, 2)
+        self.lstm.flatten_parameters()
+        x, _ = self.lstm(x)
+        return x
+    
+    def length_to_mask(self, lengths):
+        mask = torch.arange(lengths.max()).unsqueeze(0).expand(lengths.shape[0], -1).type_as(lengths)
+        mask = torch.gt(mask+1, lengths.unsqueeze(1))
+        return mask
+
+
+class UpSample1d(nn.Module):
+    def __init__(self, layer_type):
+        super().__init__()
+        self.layer_type = layer_type
+
+    def forward(self, x):
+        if self.layer_type == 'none':
+            return x
+        else:
+            return F.interpolate(x, scale_factor=2, mode='nearest')
+
+class AdainResBlk1d(nn.Module):
+    def __init__(self, dim_in, dim_out, style_dim=64, actv=nn.LeakyReLU(0.2),
+                 upsample='none', dropout_p=0.0):
+        super().__init__()
+        self.actv = actv
+        self.upsample_type = upsample
+        self.upsample = UpSample1d(upsample)
+        self.learned_sc = dim_in != dim_out
+        self._build_weights(dim_in, dim_out, style_dim)
+        self.dropout = nn.Dropout(dropout_p)
+        
+        if upsample == 'none':
+            self.pool = nn.Identity()
+        else:
+            self.pool = weight_norm(nn.ConvTranspose1d(dim_in, dim_in, kernel_size=3, stride=2, groups=dim_in, padding=1, output_padding=1))
+        
+        
+    def _build_weights(self, dim_in, dim_out, style_dim):
+        self.conv1 = weight_norm(nn.Conv1d(dim_in, dim_out, 3, 1, 1))
+        self.conv2 = weight_norm(nn.Conv1d(dim_out, dim_out, 3, 1, 1))
+        self.norm1 = AdaIN1d(style_dim, dim_in)
+        self.norm2 = AdaIN1d(style_dim, dim_out)
+        if self.learned_sc:
+            self.conv1x1 = weight_norm(nn.Conv1d(dim_in, dim_out, 1, 1, 0, bias=False))
+
+    def _shortcut(self, x):
+        x = self.upsample(x)
+        if self.learned_sc:
+            x = self.conv1x1(x)
+        return x
+
+    def _residual(self, x, s):
+        x = self.norm1(x, s)
+        x = self.actv(x)
+        x = self.pool(x)
+        x = self.conv1(self.dropout(x))
+        x = self.norm2(x, s)
+        x = self.actv(x)
+        x = self.conv2(self.dropout(x))
+        return x
+
+    def forward(self, x, s):
+        out = self._residual(x, s)
+        out = (out + self._shortcut(x)) / np.sqrt(2)
+        return out
+    
+class AdaLayerNorm(nn.Module):
+    def __init__(self, style_dim, channels, eps=1e-5):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.fc = nn.Linear(style_dim, channels*2)
+
+    def forward(self, x, s):
+        x = x.transpose(-1, -2)
+        x = x.transpose(1, -1)
+                
+        h = self.fc(s)
+        h = h.view(h.size(0), h.size(1), 1)
+        gamma, beta = torch.chunk(h, chunks=2, dim=1)
+        gamma, beta = gamma.transpose(1, -1), beta.transpose(1, -1)
+        
+        
+        x = F.layer_norm(x, (self.channels,), eps=self.eps)
+        x = (1 + gamma) * x + beta
+        return x.transpose(1, -1).transpose(-1, -2)
+
+class ProsodyPredictor(nn.Module):
+
+    def __init__(self, style_dim, d_hid, nlayers, max_dur=50, dropout=0.1):
+        super().__init__() 
+        
+        self.text_encoder = DurationEncoder(sty_dim=style_dim, 
+                                            d_model=d_hid,
+                                            nlayers=nlayers, 
+                                            dropout=dropout)
+
+        self.lstm = nn.LSTM(d_hid + style_dim, d_hid // 2, 1, batch_first=True, bidirectional=True)
+        self.duration_proj = LinearNorm(d_hid, max_dur)
+        
+        self.shared = nn.LSTM(d_hid + style_dim, d_hid // 2, 1, batch_first=True, bidirectional=True)
+        self.F0 = nn.ModuleList()
+        self.F0.append(AdainResBlk1d(d_hid, d_hid, style_dim, dropout_p=dropout))
+        self.F0.append(AdainResBlk1d(d_hid, d_hid // 2, style_dim, upsample=True, dropout_p=dropout))
+        self.F0.append(AdainResBlk1d(d_hid // 2, d_hid // 2, style_dim, dropout_p=dropout))
+
+        self.N = nn.ModuleList()
+        self.N.append(AdainResBlk1d(d_hid, d_hid, style_dim, dropout_p=dropout))
+        self.N.append(AdainResBlk1d(d_hid, d_hid // 2, style_dim, upsample=True, dropout_p=dropout))
+        self.N.append(AdainResBlk1d(d_hid // 2, d_hid // 2, style_dim, dropout_p=dropout))
+        
+        self.F0_proj = nn.Conv1d(d_hid // 2, 1, 1, 1, 0)
+        self.N_proj = nn.Conv1d(d_hid // 2, 1, 1, 1, 0)
+
+
+    def forward(self, texts, style, text_lengths, alignment, m):
+        d = self.text_encoder(texts, style, text_lengths, m)
+        
+        batch_size = d.shape[0]
+        text_size = d.shape[1]
+        
+        # predict duration
+        input_lengths = text_lengths.cpu().numpy()
+        x = nn.utils.rnn.pack_padded_sequence(
+            d, input_lengths, batch_first=True, enforce_sorted=False)
+        
+        m = m.to(text_lengths.device).unsqueeze(1)
+        
+        self.lstm.flatten_parameters()
+        x, _ = self.lstm(x)
+        x, _ = nn.utils.rnn.pad_packed_sequence(
+            x, batch_first=True)
+        
+        x_pad = torch.zeros([x.shape[0], m.shape[-1], x.shape[-1]])
+
+        x_pad[:, :x.shape[1], :] = x
+        x = x_pad.to(x.device)
+                
+        duration = self.duration_proj(nn.functional.dropout(x, 0.5, training=self.training))
+        
+        en = (d.transpose(-1, -2) @ alignment)
+
+        return duration.squeeze(-1), en
+    
+    def F0Ntrain(self, x, s):
+        x, _ = self.shared(x.transpose(-1, -2))
+        
+        F0 = x.transpose(-1, -2)
+        for block in self.F0:
+            F0 = block(F0, s)
+        F0 = self.F0_proj(F0)
+
+        N = x.transpose(-1, -2)
+        for block in self.N:
+            N = block(N, s)
+        N = self.N_proj(N)
+        
+        return F0.squeeze(1), N.squeeze(1)
+    
+    def length_to_mask(self, lengths):
+        mask = torch.arange(lengths.max()).unsqueeze(0).expand(lengths.shape[0], -1).type_as(lengths)
+        mask = torch.gt(mask+1, lengths.unsqueeze(1))
+        return mask
+
+class DurationEncoder(nn.Module):
+
+    def __init__(self, sty_dim, d_model, nlayers, dropout=0.1):
+        super().__init__()
+        self.lstms = nn.ModuleList()
+        for _ in range(nlayers):
+            self.lstms.append(nn.LSTM(d_model + sty_dim, 
+                                 d_model // 2, 
+                                 num_layers=1, 
+                                 batch_first=True, 
+                                 bidirectional=True, 
+                                 dropout=dropout))
+            self.lstms.append(AdaLayerNorm(sty_dim, d_model))
+        
+        
+        self.dropout = dropout
+        self.d_model = d_model
+        self.sty_dim = sty_dim
+
+    def forward(self, x, style, text_lengths, m):
+        masks = m.to(text_lengths.device)
+        
+        x = x.permute(2, 0, 1)
+        s = style.expand(x.shape[0], x.shape[1], -1)
+        x = torch.cat([x, s], axis=-1)
+        x.masked_fill_(masks.unsqueeze(-1).transpose(0, 1), 0.0)
+                
+        x = x.transpose(0, 1)
+        input_lengths = text_lengths.cpu().numpy()
+        x = x.transpose(-1, -2)
+        
+        for block in self.lstms:
+            if isinstance(block, AdaLayerNorm):
+                x = block(x.transpose(-1, -2), style).transpose(-1, -2)
+                x = torch.cat([x, s.permute(1, -1, 0)], axis=1)
+                x.masked_fill_(masks.unsqueeze(-1).transpose(-1, -2), 0.0)
+            else:
+                x = x.transpose(-1, -2)
+                x = nn.utils.rnn.pack_padded_sequence(
+                    x, input_lengths, batch_first=True, enforce_sorted=False)
+                block.flatten_parameters()
+                x, _ = block(x)
+                x, _ = nn.utils.rnn.pad_packed_sequence(
+                    x, batch_first=True)
+                x = F.dropout(x, p=self.dropout, training=self.training)
+                x = x.transpose(-1, -2)
+                
+                x_pad = torch.zeros([x.shape[0], x.shape[1], m.shape[-1]])
+
+                x_pad[:, :, :x.shape[-1]] = x
+                x = x_pad.to(x.device)
+        
+        return x.transpose(-1, -2)
+    
+    def inference(self, x, style):
+        x = self.embedding(x.transpose(-1, -2)) * np.sqrt(self.d_model)
+        style = style.expand(x.shape[0], x.shape[1], -1)
+        x = torch.cat([x, style], axis=-1)
+        src = self.pos_encoder(x)
+        output = self.transformer_encoder(src).transpose(0, 1)
+        return output
+    
+    def length_to_mask(self, lengths):
+        mask = torch.arange(lengths.max()).unsqueeze(0).expand(lengths.shape[0], -1).type_as(lengths)
+        mask = torch.gt(mask+1, lengths.unsqueeze(1))
+        return mask
+
+# https://github.com/yl4579/StyleTTS2/blob/main/utils.py
+def recursive_munch(d):
+    if isinstance(d, dict):
+        return Munch((k, recursive_munch(v)) for k, v in d.items())
+    elif isinstance(d, list):
+        return [recursive_munch(v) for v in d]
+    else:
+        return d
+
+def build_model(path, device):
+    config = Path(__file__).parent / 'config.json'
+    assert config.exists(), f'Config path incorrect: config.json not found at {config}'
+    with open(config, 'r') as r:
+        args = recursive_munch(json.load(r))
+    assert args.decoder.type == 'istftnet', f'Unknown decoder type: {args.decoder.type}'
+    decoder = Decoder(dim_in=args.hidden_dim, style_dim=args.style_dim, dim_out=args.n_mels,
+            resblock_kernel_sizes = args.decoder.resblock_kernel_sizes,
+            upsample_rates = args.decoder.upsample_rates,
+            upsample_initial_channel=args.decoder.upsample_initial_channel,
+            resblock_dilation_sizes=args.decoder.resblock_dilation_sizes,
+            upsample_kernel_sizes=args.decoder.upsample_kernel_sizes,
+            gen_istft_n_fft=args.decoder.gen_istft_n_fft, gen_istft_hop_size=args.decoder.gen_istft_hop_size)
+    text_encoder = TextEncoder(channels=args.hidden_dim, kernel_size=5, depth=args.n_layer, n_symbols=args.n_token)
+    predictor = ProsodyPredictor(style_dim=args.style_dim, d_hid=args.hidden_dim, nlayers=args.n_layer, max_dur=args.max_dur, dropout=args.dropout)
+    bert = load_plbert()
+    bert_encoder = nn.Linear(bert.config.hidden_size, args.hidden_dim)
+    for parent in [bert, bert_encoder, predictor, decoder, text_encoder]:
+        for child in parent.children():
+            if isinstance(child, nn.RNNBase):
+                child.flatten_parameters()
+    model = Munch(
+        bert=bert.to(device).eval(),
+        bert_encoder=bert_encoder.to(device).eval(),
+        predictor=predictor.to(device).eval(),
+        decoder=decoder.to(device).eval(),
+        text_encoder=text_encoder.to(device).eval(),
+    )
+    for key, state_dict in torch.load(path, map_location='cpu', weights_only=True)['net'].items():
+        assert key in model, key
+        try:
+            model[key].load_state_dict(state_dict)
+        except:
+            state_dict = {k[7:]: v for k, v in state_dict.items()}
+            model[key].load_state_dict(state_dict, strict=False)
+    return model

plbert.py

@@ -0,0 +1,15 @@
+# https://github.com/yl4579/StyleTTS2/blob/main/Utils/PLBERT/util.py
+from transformers import AlbertConfig, AlbertModel
+
+class CustomAlbert(AlbertModel):
+    def forward(self, *args, **kwargs):
+        # Call the original forward method
+        outputs = super().forward(*args, **kwargs)
+        # Only return the last_hidden_state
+        return outputs.last_hidden_state
+
+def load_plbert():
+    plbert_config = {'vocab_size': 178, 'hidden_size': 768, 'num_attention_heads': 12, 'intermediate_size': 2048, 'max_position_embeddings': 512, 'num_hidden_layers': 12, 'dropout': 0.1}
+    albert_base_configuration = AlbertConfig(**plbert_config)
+    bert = CustomAlbert(albert_base_configuration)
+    return bert

voices/af.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fad4192fd8a840f925b0e3fc2be54e20531f91a9ac816a485b7992ca0bd83ebf
+size 524355

voices/af_bella.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2828c6c2f94275ef3441a2edfcf48293298ee0f9b56ce70fb2e344345487b922
+size 524449

voices/af_nicole.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9401802fb0b7080c324dec1a75d60f31d977ced600a99160e095dbc5a1172692
+size 524454

voices/af_sarah.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba7918c4ace6ace4221e7e01eb3a6d16596cba9729850551c758cd2ad3a4cd08
+size 524449

voices/af_sky.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f16f1bb778de36a177ae4b0b6f1e59783d5f4d3bcecf752c3e1ee98299b335e
+size 524375

voices/am_adam.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1921528b400a553f66528c27899d95780918fe33b1ac7e2a871f6a0de475f176
+size 524444

voices/am_michael.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a255c9562c363103adc56c09b7daf837139d3bdaa8bd4dd74847ab1e3e8c28be
+size 524459

voices/bf_emma.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:992e6d8491b8926ef4a16205250e51a21d9924405a5d37e2db6e94adfd965c3b
+size 524365

voices/bf_isabella.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0865a03931230100167f7a81d394b143c072efe2d7e4c4a87b5c54d6283f580
+size 524365

voices/bm_george.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d763dfe13e934357f4d8322b718787d79e32f2181e29ca0cf6aa637d8092b96
+size 524464

voices/bm_lewis.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f70d9ea4d65f522f224628f06d86ea74279faae23bd7e765848a374aba916b76
+size 524449