app.py

import os, torchaudio, torch, gradio as gr
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from pyctcdecode import build_ctcdecoder
from huggingface_hub import hf_hub_download
from pyannote.audio import Pipeline

MODEL_ID = "vikram-iitm/indic-asr-hi-kenlm4"
HF_TOKEN = os.getenv("HF_TOKEN")

processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID).eval()
lm_path = hf_hub_download(repo_id=MODEL_ID, filename="lm.binary")
labels = [t for t,_ in sorted(processor.tokenizer.get_vocab().items(), key=lambda x: 
x[1])]
decoder = build_ctcdecoder(labels, kenlm_model_path=lm_path)
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.0", use_auth_token=HF_TOKEN)

def diarize_asr(path):
    diar = pipeline(path)
    wav, sr = torchaudio.load(path)
    if sr != 16000:
        wav = torchaudio.functional.resample(wav, sr, 16000)
    out = []
    for turn, _, speaker in diar.itertracks(yield_label=True):
        s = int(turn.start * 16000)
        e = int(turn.end * 16000)
        seg = wav[:, s:e]
        iv = processor(seg.squeeze(), sampling_rate=16000, return_tensors="pt", padding=True)
        with torch.no_grad():
            logits = model(**iv).logits[0].cpu().numpy()
        text = decoder.decode(logits.argmax(-1))
        out.append(f"[{turn.start:.1f}-{turn.end:.1f}] Speaker {speaker}: {text}")
    return "\n".join(out)

gr.Interface(
    diarize_asr,
    inputs=gr.Audio(sources=["microphone","upload"], type="filepath"),
    outputs="text",
    title="Hindi ASR + Speaker Diarization"
).launch()