app.py

import torch
import torchaudio
import gradio as gr
import time
import numpy as np
import scipy.io.wavfile
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

# ✅ 1️⃣ Use "whisper-medium" for the best balance of speed & accuracy
device = "cpu"
torch_dtype = torch.float32
MODEL_NAME = "openai/whisper-medium"

# ✅ 2️⃣ Load Whisper Model on CPU
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    MODEL_NAME, torch_dtype=torch_dtype, use_safetensors=True
)
model.to(device)

# ✅ 3️⃣ Speed up execution with torch.compile()
model = torch.compile(model)  # ✅ Faster inference on CPU

# ✅ 4️⃣ Load Processor & Pipeline
processor = AutoProcessor.from_pretrained(MODEL_NAME)
processor.feature_extractor.sampling_rate = 16000  # ✅ Set correct sampling rate

pipe = pipeline(
    task="automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    chunk_length_s=10,  # ✅ Longer chunks for better accuracy
    torch_dtype=torch_dtype,
    device=device,
    generate_kwargs={"num_beams": 5, "language": "en", "temperature": 0.1},  # ✅ Beam search + English
)

# ✅ 5️⃣ Real-Time Streaming Transcription (Microphone)
def stream_transcribe(stream, new_chunk):
    start_time = time.time()
    try:
        sr, y = new_chunk

        # ✅ Convert stereo to mono
        if y.ndim > 1:
            y = y.mean(axis=1)

        y = y.astype(np.float32)
        y /= np.max(np.abs(y))

        # ✅ Resample audio to 16kHz using torchaudio
        y_tensor = torch.tensor(y)
        y_resampled = torchaudio.functional.resample(y_tensor, orig_freq=sr, new_freq=16000).numpy()

        # ✅ Append to Stream
        if stream is not None:
            stream = np.concatenate([stream, y_resampled])
        else:
            stream = y_resampled
            
        # ✅ Run Transcription with Optimized Parameters
        transcription = pipe({"sampling_rate": 16000, "raw": stream})["text"]
        latency = time.time() - start_time

        return stream, transcription, f"{latency:.2f} sec"

    except Exception as e:
        print(f"Error: {e}")
        return stream, str(e), "Error"

# ✅ 6️⃣ Transcription for File Upload
def transcribe(inputs, previous_transcription):
    start_time = time.time()
    try:
        # ✅ Convert file input to correct format
        sample_rate, audio_data = inputs

        # ✅ Resample using torchaudio (optimized)
        audio_tensor = torch.tensor(audio_data)
        resampled_audio = torchaudio.functional.resample(audio_tensor, orig_freq=sample_rate, new_freq=16000).numpy()

        transcription = pipe({"sampling_rate": 16000, "raw": resampled_audio})["text"]

        previous_transcription += transcription
        latency = time.time() - start_time

        return previous_transcription, f"{latency:.2f} sec"

    except Exception as e:
        print(f"Error: {e}")
        return previous_transcription, "Error"

# ✅ 7️⃣ Clear Function
def clear():
    return ""

# ✅ 8️⃣ Gradio Interface (Microphone Streaming)
with gr.Blocks() as microphone:
    gr.Markdown(f"# Whisper Medium - High Accuracy Transcription (Optimized CPU) 🎙️")
    gr.Markdown(f"Using [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) for best speech-to-text performance.")

    with gr.Row():
        input_audio_microphone = gr.Audio(sources=["microphone"], type="numpy", streaming=True)
        output = gr.Textbox(label="Live Transcription", value="")
        latency_textbox = gr.Textbox(label="Latency (seconds)", value="0.0")

    with gr.Row():
        clear_button = gr.Button("Clear Output")

    state = gr.State()
    input_audio_microphone.stream(
        stream_transcribe, [state, input_audio_microphone], 
        [state, output, latency_textbox], time_limit=30, stream_every=1
    )
    clear_button.click(clear, outputs=[output])

# ✅ 9️⃣ Gradio Interface (File Upload)
with gr.Blocks() as file:
    gr.Markdown(f"# Upload Audio File for Transcription 🎵")
    gr.Markdown(f"Using [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) for better transcription accuracy.")

    with gr.Row():
        input_audio = gr.Audio(sources=["upload"], type="numpy")
        output = gr.Textbox(label="Transcription", value="")
        latency_textbox = gr.Textbox(label="Latency (seconds)", value="0.0")

    with gr.Row():
        submit_button = gr.Button("Submit")
        clear_button = gr.Button("Clear Output")

    submit_button.click(transcribe, [input_audio, output], [output, latency_textbox])
    clear_button.click(clear, outputs=[output])

# ✅ 🔟 Final Gradio App
with gr.Blocks(theme=gr.themes.Ocean()) as demo:
    gr.TabbedInterface([microphone, file], ["Microphone", "Upload Audio"])

# ✅ 1️⃣1️⃣ Run Gradio Locally
if __name__ == "__main__":
    demo.launch()