implement audio translation to danish speech

app.py

@@ -1,72 +1,114 @@
 import gradio as gr
 import numpy as np
 import torch
-from datasets import load_dataset
 
-from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline
+from transformers import pipeline
 
+checkpoint_finetuned = "JackismyShephard/speecht5_tts-finetuned-nst-da"
 
+revision = "5af228df418092b681cf31c31e413bdd2b5f9c8c"
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 # load speech translation checkpoint
-asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
+asr_pipe = pipeline(
+    "automatic-speech-recognition",
+    model="openai/whisper-base",
+    device=device,
+)
 
 # load text-to-speech checkpoint and speaker embeddings
-processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+pipe = pipeline(
+    "text-to-speech",
+    model=checkpoint_finetuned,
+    use_fast=True,
+    device=device,
+    revision=revision,
+)
 
-model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
-vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
+speaker_embedding_path = "female_23_vestjylland.npy"
+speaker_embedding = np.load(speaker_embedding_path)
+speaker_embedding_tensor = torch.tensor(speaker_embedding).unsqueeze(0)
 
-embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
-speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+target_dtype = np.int16
+max_range = np.iinfo(target_dtype).max
 
 
 def translate(audio):
-    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
+    outputs = asr_pipe(
+        audio,
+        max_new_tokens=256,
+        batch_size=8,
+        chunk_length_s=30,
+        generate_kwargs={"task": "translate", "language": "danish"},
+    )
     return outputs["text"]
 
 
 def synthesise(text):
-    inputs = processor(text=text, return_tensors="pt")
-    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
-    return speech.cpu()
+    if len(text.strip()) == 0:
+        return (16000, np.zeros(0))
+
+    text = replace_danish_letters(text)
+
+    forward_params = {"speaker_embeddings": speaker_embedding}
+    speech = pipe(text, forward_params=forward_params)
+
+    sr, audio = speech["sampling_rate"], speech["audio"]
+
+    audio = (audio * max_range).astype(np.int16)
+
+    return sr, audio
 
 
 def speech_to_speech_translation(audio):
     translated_text = translate(audio)
-    synthesised_speech = synthesise(translated_text)
-    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
-    return 16000, synthesised_speech
+    return synthesise(translated_text)
+
+
+def replace_danish_letters(text):
+    for src, dst in replacements:
+        text = text.replace(src, dst)
+    return text
+
+
+replacements = [
+    ("&", "og"),
+    ("\r", " "),
+    ("´", ""),
+    ("\\", ""),
+    ("¨", " "),
+    ("Å", "AA"),
+    ("Æ", "AE"),
+    ("É", "E"),
+    ("Ö", "OE"),
+    ("Ø", "OE"),
+    ("á", "a"),
+    ("ä", "ae"),
+    ("å", "aa"),
+    ("è", "e"),
+    ("î", "i"),
+    ("ô", "oe"),
+    ("ö", "oe"),
+    ("ø", "oe"),
+    ("ü", "y"),
+]
 
 
 title = "Cascaded STST"
 description = """
-Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in English. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
-[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:
+Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in Danish. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and JackismyShephard's
+[speecht5_tts-finetuned-nst-da](https://huggingface.co/JackismyShephard/speecht5_tts-finetuned-nst-da) model for text-to-speech:
 
 ![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
 """
 
-demo = gr.Blocks()
-
-mic_translate = gr.Interface(
+demo = gr.Interface(
     fn=speech_to_speech_translation,
-    inputs=gr.Audio(source="microphone", type="filepath"),
-    outputs=gr.Audio(label="Generated Speech", type="numpy"),
-    title=title,
-    description=description,
-)
-
-file_translate = gr.Interface(
-    fn=speech_to_speech_translation,
-    inputs=gr.Audio(source="upload", type="filepath"),
+    inputs=gr.Audio(type="filepath"),
     outputs=gr.Audio(label="Generated Speech", type="numpy"),
     examples=[["./example.wav"]],
     title=title,
     description=description,
 )
 
-with demo:
-    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])
-
 demo.launch()