app.py

import gradio as gr
import torchaudio
import speechbrain
from speechbrain.pretrained import EncoderClassifier, Tacotron2, HIFIGAN, ASR
import os
import soundfile as sf

# Ensure output directory exists
os.makedirs("output_audio", exist_ok=True)

# Load models
encoder = EncoderClassifier.from_hparams(source="speechbrain/spkrec-ecapa-voxceleb", savedir="models/encoder")
tacotron2 = Tacotron2.from_hparams(source="speechbrain/tts-tacotron2-ljspeech", savedir="models/tacotron2")
hifigan = HIFIGAN.from_hparams(source="speechbrain/tts-hifigan-ljspeech", savedir="models/hifigan")
asr = ASR.from_hparams(source="speechbrain/asr-transformer-librispeech", savedir="models/asr")

def speech_to_text(input_audio):
    sig, sr = torchaudio.load(input_audio)
    transcription = asr.transcribe_file(input_audio)
    return transcription

def speech_to_speech(input_audio, target_text):
    # Load and encode speaker from input audio
    signal, fs = torchaudio.load(input_audio)
    if fs != 16000:
        signal = torchaudio.transforms.Resample(orig_freq=fs, new_freq=16000)(signal)
    embedding = encoder.encode_batch(signal)

    # Synthesize speech from text
    mel_output, mel_length, alignment = tacotron2.encode_text(target_text, embedding)
    waveform = hifigan.decode_batch(mel_output)

    # Save output audio
    output_path = "output_audio/synthesized_speech.wav"
    sf.write(output_path, waveform.squeeze().cpu().numpy(), 22050)
    return output_path

def text_to_speech(text):
    mel_output, mel_length, alignment = tacotron2.encode_text(text)
    waveform = hifigan.decode_batch(mel_output)

    output_path = "output_audio/text_to_speech.wav"
    sf.write(output_path, waveform.squeeze().cpu().numpy(), 22050)
    return output_path

iface = gr.Interface(
    fn={
        "Speech to Text": speech_to_text,
        "Text to Speech": text_to_speech,
        "Speech to Speech": speech_to_speech
    },
    inputs={
        "Speech to Text": gr.inputs.Audio(source="upload", type="file"),
        "Text to Speech": gr.inputs.Textbox(label="Text"),
        "Speech to Speech": [gr.inputs.Audio(source="upload", type="file"), gr.inputs.Textbox(label="Target Text")]
    },
    outputs={
        "Speech to Text": gr.outputs.Textbox(label="Transcription"),
        "Text to Speech": gr.outputs.Audio(type="file", label="Synthesized Speech"),
        "Speech to Speech": gr.outputs.Audio(type="file", label="Synthesized Speech")
    },
    title="Speech Processing App",
    description="Upload an audio file or enter text to perform various speech processing tasks.",
    layout="vertical"
)

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