sprachText/stt_test.py

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
import sounddevice as sd
import numpy as np
import wave
import gradio as gr
import time
from pydub import AudioSegment
import os

# Setup device
device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

# Load Whisper model
model_id = "openai/whisper-large-v3-turbo"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

# Audio recording settings
SAMPLE_RATE = 16000  # Whisper prefers 16 kHz
FILENAME = "recorded_audio.wav"
is_recording = False
start_time = None  # Track the recording start time
recorded_audio = None


def start_recording():
    """Starts recording audio."""
    global is_recording, recorded_audio, start_time
    is_recording = True
    start_time = time.time()  # Record the start time
    print("Recording started...")
    recorded_audio = sd.rec(int(SAMPLE_RATE * 60), samplerate=SAMPLE_RATE, channels=1, dtype=np.float32)
    return "Recording... Click 'Stop Recording' to finish."


def stop_recording():
    """Stops recording audio and saves it."""
    global is_recording, recorded_audio, start_time
    if not is_recording:
        return "Not recording!"
    sd.stop()
    is_recording = False
    elapsed_time = time.time() - start_time  # Calculate elapsed time
    print(f"Recording stopped. Duration: {elapsed_time:.2f} seconds.")
    save_audio_to_wav(recorded_audio[:int(SAMPLE_RATE * elapsed_time)], FILENAME)  # Truncate to actual duration
    return "Recording stopped. Click 'Transcribe' to see the result."


def save_audio_to_wav(audio, filename):
    """Saves audio data to a WAV file."""
    audio = (audio * 32767).astype(np.int16)  # Convert to 16-bit PCM format
    with wave.open(filename, 'w') as wf:
        wf.setnchannels(1)  # Mono
        wf.setsampwidth(2)  # 2 bytes per sample
        wf.setframerate(SAMPLE_RATE)
        wf.writeframes(audio.tobytes())

def get_audio_duration(filename):
    """Returns the duration of the audio file in seconds."""
    audio = AudioSegment.from_wav(filename)
    return len(audio) / 1000

def split_audio(filename, chunk_length_ms=30000):
    """Splits an audio file into chunks."""
    audio = AudioSegment.from_wav(filename)
    chunks = [audio[i:i + chunk_length_ms] for i in range(0, len(audio), chunk_length_ms)]
    return chunks


def transcribe_audio():
    """Transcribes the audio file using Whisper."""
    print("Checking audio duration...")
    duration = get_audio_duration(FILENAME)

    if duration > 30:
        print(f"Audio is too long ({duration:.2f} seconds). Splitting into chunks...")
        chunks = split_audio(FILENAME)
        transcription = []

        for i, chunk in enumerate(chunks):
            chunk_filename = f"chunk_{i}.wav"
            chunk.export(chunk_filename, format="wav")
            print(f"Transcribing chunk {i + 1}/{len(chunks)}...")
            result = pipe(chunk_filename)
            transcription.append(result["text"])

            # Clean up temporary chunk file
            os.remove(chunk_filename)

        print("Transcription complete.")
        return " ".join(transcription)
    else:
        print(f"Audio is short enough ({duration:.2f} seconds). Transcribing directly...")
        result = pipe(FILENAME)
        print("Transcription complete.")
        return result["text"]


# Gradio Interface
with gr.Blocks() as interface:
    gr.Markdown("# Voice to Text App")
    gr.Markdown("Click 'Start Recording' to record your voice, 'Stop Recording' to save, and 'Transcribe' to convert speech to text.")

    start_button = gr.Button("Start Recording")
    stop_button = gr.Button("Stop Recording")
    transcribe_button = gr.Button("Transcribe")

    output = gr.Textbox(label="Output")

    start_button.click(start_recording, outputs=output)
    stop_button.click(stop_recording, outputs=output)
    transcribe_button.click(transcribe_audio, outputs=output)

if __name__ == "__main__":
    interface.launch()