app.py

import gradio as gr
import torch
import zipfile
from pyctcdecode import build_ctcdecoder
from speechbrain.pretrained import EncoderASR
from transformers.file_utils import cached_path, hf_bucket_url


def download_lm(cache_dir="./cache"):
    cache_dir = "./cache/"
    lm_file = hf_bucket_url("dragonSwing/wav2vec2-base-vn-270h", filename="4gram.zip")
    lm_file = cached_path(lm_file, cache_dir=cache_dir)
    with zipfile.ZipFile(lm_file, "r") as zip_ref:
        zip_ref.extractall(cache_dir)
    lm_file = cache_dir + "lm.binary"
    vocab_file = cache_dir + "vocab-260000.txt"
    return lm_file, vocab_file


model = EncoderASR.from_hparams(
    source="dragonSwing/wav2vec2-base-vn-270h", savedir="./pretrained/wav2vec-vi-asr"
)


def get_decoder_ngram_model(tokenizer, ngram_lm_path, vocab_path=None):
    unigrams = None
    if vocab_path is not None:
        unigrams = []
        with open(vocab_path, encoding="utf-8") as f:
            for line in f:
                unigrams.append(line.strip())

    vocab_dict = tokenizer.get_vocab()
    sort_vocab = sorted((value, key) for (key, value) in vocab_dict.items())
    vocab = [x[1] for x in sort_vocab]
    vocab_list = vocab

    # convert ctc blank character representation
    vocab_list[tokenizer.pad_token_id] = ""
    # replace special characters
    vocab_list[tokenizer.word_delimiter_token_id] = " "
    # specify ctc blank char index, since conventially it is the last entry of the logit matrix
    decoder = build_ctcdecoder(vocab_list, ngram_lm_path, unigrams=unigrams)
    return decoder


# ngram_lm_model = get_decoder_ngram_model(model.tokenizer, lm_file, vocab_file)


def transcribe_file(path, max_seconds=20, lm_model=None):
    waveform = model.load_audio(path)
    if max_seconds > 0:
        waveform = waveform[: max_seconds * 16000]
    batch = waveform.unsqueeze(0)
    rel_length = torch.tensor([1.0])
    if lm_model:
        with torch.no_grad():
            logits = model(batch, rel_length)
        text_batch = [
            lm_model.decode(logit.detach().cpu().numpy(), beam_width=500)
            for logit in logits
        ]
        return text_batch[0]
    else:
        text_batch, _ = model.transcribe_batch(
            batch, rel_length
        )
        return text_batch[0]

def speech_recognize(file_upload, file_mic):
    if file_upload is not None:
        file = file_upload
    elif file_mic is not None:
        file = file_mic
    else:
        return ""
    # text = model.transcribe_file(file)
    text = transcribe_file(file)
    return text


inputs = [
    gr.Audio(source="upload", type="filepath", optional=True),
    gr.Audio(source="microphone", type="filepath", optional=True),
]
outputs = gr.Textbox(label="Output Text")
title = "wav2vec2-base-vietnamese-270h"
description = "Gradio demo for a wav2vec2 base vietnamese speech recognition. To use it, simply upload your audio, click one of the examples to load them, or record from your own microphone. Read more at the links below. Currently supports 16_000hz audio files"
article = "<p style='text-align: center'><a href='https://huggingface.co./dragonSwing/wav2vec2-base-vn-270h' target='_blank'>Pretrained model</a></p>"
examples = [
    ["example1.wav", "example1.wav"],
    ["example2.mp3", "example2.mp3"],
    ["example3.mp3", "example3.mp3"],
    ["example4.wav", "example4.wav"],
]
gr.Interface(
    speech_recognize,
    inputs,
    outputs,
    title=title,
    description=description,
    article=article,
    examples=examples,
).launch()