from TTS.api import TTS from pydub import AudioSegment import os import re import ffmpeg import shutil import argparse import torch os.environ["COQUI_TOS_AGREED"] = "1" def adjust_speed(input_file, speed_factor): output_file = input_file.replace(".wav", "_adjusted.wav") ffmpeg.input(input_file).filter('atempo', speed_factor).output(output_file, acodec='pcm_s16le').run() return output_file def generate_speech(text, speaker_voice_map, output_file): combined_audio = AudioSegment.empty() temp_files = [] if torch.cuda.is_available(): tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda") else: tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2") for line in text.split("\n"): if not line.strip(): continue match = re.match(r"\[SPEAKER_(\d+)\] \[(\d+\.\d+)-(\d+\.\d+)\] (.+)", line) if not match: continue speaker_id, start_time, end_time, sentence = match.groups() start_time, end_time = float(start_time), float(end_time) segment_duration = (end_time - start_time) * 1000 # Duration in milliseconds speaker_wav = speaker_voice_map.get(f"SPEAKER_{speaker_id}") if not speaker_wav: continue os.makedirs('./audio/temp', exist_ok=True) temp_file_path = f"./audio/temp/temp_output_part_{len(temp_files)}.wav" temp_files.append(temp_file_path) tts_speed = 1.0 tts.tts_to_file(text=sentence, file_path=temp_file_path, speaker_wav=speaker_wav, language="es", speed=tts_speed) segment_audio = AudioSegment.from_wav(temp_file_path) if segment_audio.duration_seconds * 1000 > segment_duration: while tts_speed < 2.0 and segment_audio.duration_seconds * 1000 > segment_duration: tts_speed += 0.5 tts.tts_to_file(text=sentence, file_path=temp_file_path, speaker_wav=speaker_wav, language="es", speed=tts_speed) segment_audio = AudioSegment.from_wav(temp_file_path) if segment_audio.duration_seconds * 1000 > segment_duration: required_speed = segment_duration / (segment_audio.duration_seconds * 1000) if required_speed < 1.0: required_speed = 1.0 / required_speed temp_file_path = adjust_speed(temp_file_path, required_speed) segment_audio = AudioSegment.from_wav(temp_file_path) if combined_audio.duration_seconds == 0 and start_time > 0: combined_audio = AudioSegment.silent(duration=start_time * 1000) + combined_audio if segment_audio.duration_seconds * 1000 > segment_duration: segment_audio = segment_audio[:segment_duration] else: segment_audio = segment_audio + AudioSegment.silent(duration=segment_duration - len(segment_audio)) combined_audio += segment_audio combined_audio.export(output_file, format="wav") for temp_file in temp_files: os.remove(temp_file) def map_speaker_ids(directory): speaker_voice_map = {} for file in os.listdir(directory): if file.endswith(".wav"): speaker_id = file.replace(".wav", "") speaker_voice_map[speaker_id] = os.path.join(directory, file) return speaker_voice_map def main(speaker_directory, aligned_text_file, output_audio_file): speaker_voice_map = map_speaker_ids(speaker_directory) with open(aligned_text_file, 'r') as file: translated_text = file.read() generate_speech(translated_text, speaker_voice_map, output_audio_file) if os.path.exists('./audio/temp'): shutil.rmtree('./audio/temp') if __name__ == "__main__": parser = argparse.ArgumentParser(description="Generate speech from translated text") parser.add_argument("speaker_directory", help="Directory containing speaker voice clips") parser.add_argument("aligned_text_file", help="Path to the translated and aligned text file") parser.add_argument("output_audio_file", help="Path to save the generated speech audio file") args = parser.parse_args() main(args.speaker_directory, args.aligned_text_file, args.output_audio_file)