lid.py

from transformers import Wav2Vec2ForSequenceClassification, AutoFeatureExtractor
import torch
import librosa
import numpy as np
import os

# Load Facebook MMS Language Identification Model
MODEL_ID = "facebook/mms-lid-1024"
processor = AutoFeatureExtractor.from_pretrained(MODEL_ID)
model = Wav2Vec2ForSequenceClassification.from_pretrained(MODEL_ID)

# Constants
LID_SAMPLING_RATE = 16_000
LID_THRESHOLD = 0.33  # Confidence threshold
LID_LANGUAGES = {}

# Load Language Labels
LANG_FILE = "data/lid/all_langs.tsv"
if not os.path.exists(LANG_FILE):
    raise FileNotFoundError(f"Language file '{LANG_FILE}' not found!")

with open(LANG_FILE, encoding="utf-8") as f:
    for line in f:
        iso, name = line.strip().split(" ", 1)
        LID_LANGUAGES[iso] = name

# Identify Audio Language
def identify(audio_data=None):
    if not audio_data:
        return "<<ERROR: Empty Audio Input>>"
    
    # Microphone Input
    if isinstance(audio_data, tuple):
        sr, audio_samples = audio_data
        audio_samples = (audio_samples / 32768.0).astype(np.float32)
        if sr != LID_SAMPLING_RATE:
            audio_samples = librosa.resample(audio_samples, orig_sr=sr, target_sr=LID_SAMPLING_RATE)

    # File Upload
    elif isinstance(audio_data, str):
        if not os.path.exists(audio_data):
            return f"<<ERROR: File '{audio_data}' not found!>>"
        audio_samples, _ = librosa.load(audio_data, sr=LID_SAMPLING_RATE, mono=True)

    else:
        return "<<ERROR: Invalid Audio Input>>"

    # Process Input
    inputs = processor(audio_samples, sampling_rate=LID_SAMPLING_RATE, return_tensors="pt")

    # Select Device
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device)
    inputs = inputs.to(device)

    # Predict Language
    with torch.no_grad():
        logit = model(**inputs).logits

    # Compute Probabilities
    logit_lsm = torch.log_softmax(logit.squeeze(), dim=-1)
    scores, indices = torch.topk(logit_lsm, 5, dim=-1)
    scores, indices = torch.exp(scores).cpu().tolist(), indices.cpu().tolist()

    # Map to Language Labels
    iso2score = {model.config.id2label[int(i)]: s for s, i in zip(scores, indices)}

    # Confidence Check
    if max(iso2score.values()) < LID_THRESHOLD:
        return "Low confidence in language detection. No output shown."

    return {LID_LANGUAGES.get(iso, iso): score for iso, score in iso2score.items()}

# Example Usage
LID_EXAMPLES = [
    ["upload/english.mp3"],
    ["upload/tamil.mp3"],
    ["upload/burmese.mp3"],
]