app.py

import gradio as gr
import torch
import numpy as np
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from scipy.signal import resample

# Load model and processor
model_id = "facebook/wav2vec2-large-960h-lv60-self"
processor = Wav2Vec2Processor.from_pretrained(model_id)
model = Wav2Vec2ForCTC.from_pretrained(model_id)

# Transcription function
def transcribe(audio_tuple):
    if audio_tuple is None:
        return "⚠️ No audio received."

    audio, sample_rate = audio_tuple

    if sample_rate is None or audio is None:
        return "⚠️ Audio or sample rate missing."

    # Flatten if stereo (2D) to mono (1D)
    if len(audio.shape) == 2:
        audio = np.mean(audio, axis=1)

    # Resample if not 16000Hz
    if sample_rate != 16000:
        number_of_samples = round(len(audio) * 16000 / sample_rate)
        audio = resample(audio, number_of_samples)

    # Normalize to [-1, 1]
    audio = audio.astype(np.float32)

    # Process audio
    input_values = processor(audio, sampling_rate=16000, return_tensors="pt").input_values

    # Inference
    with torch.no_grad():
        logits = model(input_values).logits
        predicted_ids = torch.argmax(logits, dim=-1)
        transcription = processor.batch_decode(predicted_ids)[0]

    return transcription.lower()

# Gradio interface
demo = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(sources=["microphone"], type="numpy", label="🎤 Speak a word or letter"),
    outputs=gr.Textbox(label="📝 Transcription"),
    title="🔤 Wav2Vec2 Speech Transcriber",
    description="Speak into the mic and get real-time transcription using Hugging Face Wav2Vec2."
)

demo.launch()