app.py

import os
import torch
import gradio as gr
import librosa
import numpy as np
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq

# ----------------------------
# Config
# ----------------------------
ASR_MODEL_ID = "openai/whisper-large-v3"
HF_TOKEN = os.getenv("HF_TOKEN")

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32

# ----------------------------
# Load Whisper
# ----------------------------
processor = AutoProcessor.from_pretrained(
    ASR_MODEL_ID,
    use_auth_token=HF_TOKEN
)

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    ASR_MODEL_ID,
    torch_dtype=DTYPE,
    low_cpu_mem_usage=True,
    use_safetensors=True,
    use_auth_token=HF_TOKEN
).to(DEVICE)

model.eval()

# ----------------------------
# Audio preprocessing
# ----------------------------
def preprocess_audio(audio):
    if audio is None:
        return None

    sr, speech = audio

    # Stereo → mono
    if speech.ndim > 1:
        speech = np.mean(speech, axis=1)

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

    # Normalize volume
    rms = np.sqrt(np.mean(speech ** 2))
    if rms > 0:
        speech = speech / rms

    # Trim silence
    speech, _ = librosa.effects.trim(speech, top_db=25)

    # Force 16kHz
    if sr != 16000:
        speech = librosa.resample(speech, orig_sr=sr, target_sr=16000).astype(np.float32)

    return speech

# ----------------------------
# Transcription
# ----------------------------
def transcribe_audio(audio):
    speech = preprocess_audio(audio)

    if speech is None or len(speech) < 16000:
        return "Audio too short or unclear. Please speak clearly and try again."

    inputs = processor(
        speech,
        sampling_rate=16000,
        return_tensors="pt"
    )
    inputs = {k: v.to(DEVICE) for k, v in inputs.items()}

    MAX_DECODER_TOKENS = 448
    START_TOKENS = 4
    max_new_tokens = MAX_DECODER_TOKENS - START_TOKENS  # 444
    
    with torch.no_grad():
        generated_ids = model.generate(
            **inputs,
            task="transcribe",
            language="yo",
            max_new_tokens=max_new_tokens,
            temperature=0.0,
            no_repeat_ngram_size=3
        )


    text = processor.batch_decode(
        generated_ids,
        skip_special_tokens=True
    )[0].strip()

    if len(text.split()) < 2:
        return "Speech unclear. Please repeat slowly in Yoruba."

    return text

# ----------------------------
# Gradio UI
# ----------------------------
demo = gr.Interface(
    fn=transcribe_audio,
    inputs=gr.Audio(
        sources=["microphone", "upload"],
        type="numpy",
        label="Speak clearly or upload audio in Yoruba"
    ),
    outputs=gr.Textbox(label="Transcription"),
    title="Yoruba ASR (Whisper)",
    description="Speech-to-text system that transcribes only Yoruba"
)

if __name__ == "__main__":
    demo.launch(share=True)