README.md

metadata

language: fa
datasets:
  - common_voice_6_1
tags:
  - audio
  - automatic-speech-recognition
license: mit
widget:
  - example_title: Common Voice Sample 1
    src: >-
      https://datasets-server.huggingface.co/assets/common_voice/--/fa/train/0/audio/audio.mp3
  - example_title: Common Voice Sample 2
    src: >-
      https://datasets-server.huggingface.co/assets/common_voice/--/fa/train/1/audio/audio.mp3
model-index:
  - name: Sharif-wav2vec2
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice Corpus 6.1 (clean)
          type: common_voice_6_1
          config: clean
          split: test
          args:
            language: fa
        metrics:
          - name: Test WER
            type: wer
            value: 6

Sharif-wav2vec2

This is the fine-tuned version of Sharif Wav2vec2 for Farsi. The base model was fine-tuned on 108 hours of Commonvoice's Farsi samples with a sampling rate equal to 16kHz. Afterward, we trained a 5gram using kenlm toolkit and used it in the processor which increased our accuracy on online ASR.

Usage

When using the model make sure that your speech input is sampled at 16Khz. Prior to the usage, you may need to install the below dependencies:

pip install pyctcdecode
pip install pypi-kenlm

For testing you can use the hosted inference API at the hugging face (There are provided examples from common voice) it may take a while to transcribe the given voice. Or you can use the bellow code for a local run:

import tensorflow
import torchaudio
import torch
import numpy as np

from transformers import AutoProcessor, AutoModelForCTC

processor = AutoProcessor.from_pretrained("SLPL/Sharif-wav2vec2")
model = AutoModelForCTC.from_pretrained("SLPL/Sharif-wav2vec2")

speech_array, sampling_rate = torchaudio.load("path/to/your.wav")
speech_array = speech_array.squeeze().numpy()

features = processor(
    speech_array,
    sampling_rate=processor.feature_extractor.sampling_rate,
    return_tensors="pt",
    padding=True)

with torch.no_grad():
    logits = model(
        features.input_values,
        attention_mask=features.attention_mask).logits
    prediction = processor.batch_decode(logits.numpy()).text

print(prediction[0])
# تست

Evaluation

For the evaluation use the code below:

?

Result (WER):

clean	other
3.4	8.6

Citation

If you want to cite this model you can use this:

?