asr.py

import librosa
from transformers import Wav2Vec2ForCTC, AutoProcessor
import torch
import json

from huggingface_hub import hf_hub_download
from torchaudio.models.decoder import ctc_decoder

ASR_SAMPLING_RATE = 16_000

ASR_LANGUAGES = {}
with open(f"data/asr/all_langs.tsv") as f:
    for line in f:
        iso, name = line.split(" ", 1)
        ASR_LANGUAGES[iso] = name

MODEL_ID = "facebook/mms-1b-all"

processor = AutoProcessor.from_pretrained(MODEL_ID)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)


lm_decoding_config = {}
lm_decoding_configfile = hf_hub_download(
    repo_id="facebook/mms-cclms",
    filename="decoding_config.json",
    subfolder="mms-1b-all",
)

with open(lm_decoding_configfile) as f:
    lm_decoding_config = json.loads(f.read())

# allow language model decoding for specific languages
lm_decode_isos = ["eng"]


def transcribe(
    audio_source=None, microphone=None, file_upload=None, lang="eng (English)"
):
    if type(microphone) is dict:
        # HACK: microphone variable is a dict when running on examples
        microphone = microphone["name"]
    audio_fp = (
        file_upload if "upload" in str(audio_source or "").lower() else microphone
    )
    audio_samples = librosa.load(audio_fp, sr=ASR_SAMPLING_RATE, mono=True)[0]

    lang_code = lang.split()[0]
    processor.tokenizer.set_target_lang(lang_code)
    model.load_adapter(lang_code)

    inputs = processor(
        audio_samples, sampling_rate=ASR_SAMPLING_RATE, return_tensors="pt"
    )

    # set device
    if torch.cuda.is_available():
        device = torch.device("cuda")
    elif (
        hasattr(torch.backends, "mps")
        and torch.backends.mps.is_available()
        and torch.backends.mps.is_built()
    ):
        device = torch.device("mps")
    else:
        device = torch.device("cpu")

    model.to(device)
    inputs = inputs.to(device)

    with torch.no_grad():
        outputs = model(**inputs).logits

    if lang_code not in lm_decoding_config or lang_code not in lm_decode_isos:
        ids = torch.argmax(outputs, dim=-1)[0]
        transcription = processor.decode(ids)
    else:
        decoding_config = lm_decoding_config[lang_code]

        lm_file = hf_hub_download(
            repo_id="facebook/mms-cclms",
            filename=decoding_config["lmfile"].rsplit("/", 1)[1],
            subfolder=decoding_config["lmfile"].rsplit("/", 1)[0],
        )
        token_file = hf_hub_download(
            repo_id="facebook/mms-cclms",
            filename=decoding_config["tokensfile"].rsplit("/", 1)[1],
            subfolder=decoding_config["tokensfile"].rsplit("/", 1)[0],
        )
        lexicon_file = None
        if decoding_config["lexiconfile"] is not None:
            lexicon_file = hf_hub_download(
                repo_id="facebook/mms-cclms",
                filename=decoding_config["lexiconfile"].rsplit("/", 1)[1],
                subfolder=decoding_config["lexiconfile"].rsplit("/", 1)[0],
            )

        beam_search_decoder = ctc_decoder(
            lexicon=lexicon_file,
            tokens=token_file,
            lm=lm_file,
            nbest=1,
            beam_size=500,
            beam_size_token=50,
            lm_weight=float(decoding_config["lmweight"]),
            word_score=float(decoding_config["wordscore"]),
            sil_score=float(decoding_config["silweight"]),
            blank_token="<s>",
        )
        beam_search_result = beam_search_decoder(outputs.to("cpu"))
        transcription = " ".join(beam_search_result[0][0].words).strip()

    return transcription


ASR_EXAMPLES = [
    [None, "assets/english.mp3", None, "eng (English)"],
    # [None, "assets/tamil.mp3", None, "tam (Tamil)"],
    # [None, "assets/burmese.mp3", None, "mya (Burmese)"],
]

ASR_NOTE = """
The above demo uses beam-search decoding with LM for English and greedy decoding results for all other languages. 
Checkout the instructions [here](https://huggingface.co/facebook/mms-1b-all) on how to run LM decoding for other languages.
"""