app.py

import torch
import gradio as gr
from nemo.collections.asr.models import EncDecRNNTBPEModel
import soundfile as sf
import numpy as np
import torchaudio
MODEL_NAME = "ARTPARK-IISc/Vaani-FastConformer-Multilingual"

print("Loading model, this may take a few minutes...")
model = EncDecRNNTBPEModel.from_pretrained(MODEL_NAME)
model.eval()

#Use CPU if GPU is not available
if not torch.cuda.is_available():
    model = model.cpu()
print("Model loaded successfully.")
TARGET_SR = 16000

def resample_if_needed(audio, sr):
    if sr == TARGET_SR:
        return audio

    audio_tensor = torch.from_numpy(audio).unsqueeze(0)  # (1, T)

    resampler = torchaudio.transforms.Resample(
        orig_freq=sr,
        new_freq=TARGET_SR
    )

    audio_resampled = resampler(audio_tensor)
    return audio_resampled.squeeze(0).numpy()


def transcribe(audio_input):
    """
    audio_input: (sample_rate, audio_array)
    """
    if audio_input is None:
        return ""

    sr, audio = audio_input

    # Convert stereo → mono
    if audio.ndim == 2:
        audio = np.mean(audio, axis=1)

    # Convert to float32
    audio = audio.astype(np.float32)

    # Normalize
    audio = audio / (np.max(np.abs(audio)) + 1e-9)

    # Resample to 16kHz if needed
    audio = resample_if_needed(audio, sr)

    hypotheses = model.transcribe(
        audio=[audio],
        return_hypotheses=True
    )

    return hypotheses[0].text if hypotheses else ""


demo = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(
        sources=["microphone", "upload"],
        type="numpy",
        label="Record or upload WAV audio"
    ),
    outputs=gr.Textbox(label="Transcription"),
    title="Vaani Multilingual ASR (NeMo RNNT)",
    description="Upload a WAV file and get the multilingual ASR transcription."
)

demo.launch()