synthesis.py

import os
import sys
import re
import numpy as np
import torch
import soundfile as sf
import spaces
from config import models_path, results_path, sample_path, BASE_DIR
from sentence_splitter import PersianSentenceSplitter
from persian_numbers import find_and_normalize_numbers

encoder = None
synthesizer = None
vocoder = None
sentence_splitter = None

def load_models():
    global encoder, synthesizer, vocoder, sentence_splitter

    try:
        sys.path.append(os.path.join(BASE_DIR, 'pmt2'))

        from encoder import inference as encoder_module
        from synthesizer.inference import Synthesizer
        from parallel_wavegan.utils import load_model as vocoder_hifigan

        global encoder
        encoder = encoder_module

        print("Loading encoder model...")
        encoder.load_model(os.path.join(models_path, 'encoder.pt'))

        print("Loading synthesizer model...")
        synthesizer = Synthesizer(os.path.join(models_path, 'synthesizer.pt'))

        print("Loading HiFiGAN vocoder...")
        vocoder = vocoder_hifigan(os.path.join(models_path, 'vocoder_HiFiGAN.pkl'))
        vocoder.remove_weight_norm()
        vocoder = vocoder.eval().to('cuda' if torch.cuda.is_available() else 'cpu')

        sentence_splitter = PersianSentenceSplitter(max_chars=150, min_chars=30)

        print("Models loaded successfully!")
        return True
    except Exception as e:
        import traceback
        print(f"Error loading models: {traceback.format_exc()}")
        return False


def normalize_text_for_synthesis(text: str) -> str:
    text = text.replace('ك', 'ک').replace('ي', 'ی')

    text = text.replace('_', '\u200c')

    text = re.sub(r'\s+', ' ', text)
    text = text.strip()

    text = find_and_normalize_numbers(text)

    return text


def synthesize_segment(text_segment: str, embed: np.ndarray) -> np.ndarray:
    try:
        text_segment = normalize_text_for_synthesis(text_segment)

        specs = synthesizer.synthesize_spectrograms([text_segment], [embed])
        spec = specs[0]

        x = torch.from_numpy(spec.T).to('cuda' if torch.cuda.is_available() else 'cpu')

        with torch.no_grad():
            wav = vocoder.inference(x)

        wav = wav.cpu().numpy()

        if wav.ndim > 1:
            wav = wav.squeeze()

        return wav

    except Exception as e:
        import traceback
        print(f"Error synthesizing segment '{text_segment[:50]}...': {traceback.format_exc()}")
        return None


def add_silence(duration_ms: int = 300) -> np.ndarray:
    sample_rate = synthesizer.sample_rate
    num_samples = int(sample_rate * duration_ms / 1000)
    return np.zeros(num_samples, dtype=np.float32)


@spaces.GPU(duration=120)
def generate_speech(text, reference_audio=None, add_pauses: bool = True):
    if not text or text.strip() == "":
        return None

    try:
        if reference_audio is None:
            ref_wav_path = sample_path
        else:
            ref_wav_path = os.path.join(results_path, "reference_audio.wav")
            sf.write(ref_wav_path, reference_audio[1], reference_audio[0])

        print(f"Using reference audio: {ref_wav_path}")

        wav = synthesizer.load_preprocess_wav(ref_wav_path)

        encoder_wav = encoder.preprocess_wav(wav)
        embed, partial_embeds, _ = encoder.embed_utterance(encoder_wav, return_partials=True)

        text_segments = sentence_splitter.split(text)

        print(f"Split text into {len(text_segments)} segments:")
        for i, segment in enumerate(text_segments, 1):
            print(f"  Segment {i}: {segment[:60]}{'...' if len(segment) > 60 else ''}")

        audio_segments = []
        silence = add_silence(300) if add_pauses else None  # 300ms pause

        for i, segment in enumerate(text_segments):
            print(f"Processing segment {i+1}/{len(text_segments)}...")

            segment_wav = synthesize_segment(segment, embed)

            if segment_wav is not None:
                segment_wav = segment_wav.flatten() if segment_wav.ndim > 1 else segment_wav
                audio_segments.append(segment_wav)

                if add_pauses and i < len(text_segments) - 1:
                    audio_segments.append(silence)
            else:
                print(f"Warning: Failed to synthesize segment {i+1}")

        if not audio_segments:
            print("Error: No audio segments were generated successfully")
            return None

        audio_segments = [seg.flatten() if seg.ndim > 1 else seg for seg in audio_segments]

        final_wav = np.concatenate(audio_segments)

        final_wav = final_wav / np.abs(final_wav).max() * 0.97

        output_filename = f"generated_{abs(hash(text)) % 100000}.wav"
        output_path = os.path.join(results_path, output_filename)
        sf.write(output_path, final_wav, synthesizer.sample_rate)

        print(f"✓ Successfully generated speech: {output_path}")
        print(f"  Total duration: {len(final_wav) / synthesizer.sample_rate:.2f} seconds")

        return output_path

    except Exception as e:
        import traceback
        error_details = traceback.format_exc()
        print(f"Error generating speech: {error_details}")
        return None