#!/usr/bin/env python # Copyright (c) Meta Platforms, Inc. and affiliates # All rights reserved. # # This source code is licensed under the license found in the # MIT_LICENSE file in the root directory of this source tree. import os import pathlib import tempfile from pydub import AudioSegment, silence import gradio as gr import torch import torchaudio from fairseq2.assets import InProcAssetMetadataProvider, asset_store from fairseq2.data import Collater, SequenceData, VocabularyInfo from fairseq2.data.audio import ( AudioDecoder, WaveformToFbankConverter, WaveformToFbankOutput, ) from seamless_communication.inference import SequenceGeneratorOptions from fairseq2.generation import NGramRepeatBlockProcessor from fairseq2.memory import MemoryBlock from fairseq2.typing import DataType, Device from huggingface_hub import snapshot_download from seamless_communication.inference import BatchedSpeechOutput, Translator, SequenceGeneratorOptions from seamless_communication.models.generator.loader import load_pretssel_vocoder_model from seamless_communication.models.unity import ( UnitTokenizer, load_gcmvn_stats, load_unity_text_tokenizer, load_unity_unit_tokenizer, ) from torch.nn import Module from seamless_communication.cli.expressivity.evaluate.pretssel_inference_helper import PretsselGenerator from utils import LANGUAGE_CODE_TO_NAME DESCRIPTION = """\ # Seamless Expressive [SeamlessExpressive](https://github.com/facebookresearch/seamless_communication) is a speech-to-speech translation model that captures certain underexplored aspects of prosody such as speech rate and pauses, while preserving the style of one's voice and high content translation quality. """ CACHE_EXAMPLES = os.getenv("CACHE_EXAMPLES") == "1" and torch.cuda.is_available() CHECKPOINTS_PATH = pathlib.Path(os.getenv("CHECKPOINTS_PATH", "/workspace/seamless_communication/demo/expressive/models")) if not CHECKPOINTS_PATH.exists(): snapshot_download(repo_id="facebook/seamless-expressive", repo_type="model", local_dir=CHECKPOINTS_PATH) snapshot_download(repo_id="facebook/seamless-m4t-v2-large", repo_type="model", local_dir=CHECKPOINTS_PATH) # Ensure that we do not have any other environment resolvers and always return # "demo" for demo purposes. asset_store.env_resolvers.clear() asset_store.env_resolvers.append(lambda: "demo") # Construct an `InProcAssetMetadataProvider` with environment-specific metadata # that just overrides the regular metadata for "demo" environment. Note the "@demo" suffix. demo_metadata = [ { "name": "seamless_expressivity@demo", "checkpoint": f"file://{CHECKPOINTS_PATH}/m2m_expressive_unity.pt", "char_tokenizer": f"file://{CHECKPOINTS_PATH}/spm_char_lang38_tc.model", }, { "name": "vocoder_pretssel@demo", "checkpoint": f"file://{CHECKPOINTS_PATH}/pretssel_melhifigan_wm-final.pt", }, { "name": "seamlessM4T_v2_large@demo", "checkpoint": f"file://{CHECKPOINTS_PATH}/seamlessM4T_v2_large.pt", "char_tokenizer": f"file://{CHECKPOINTS_PATH}/spm_char_lang38_tc.model", }, ] asset_store.metadata_providers.append(InProcAssetMetadataProvider(demo_metadata)) LANGUAGE_NAME_TO_CODE = {v: k for k, v in LANGUAGE_CODE_TO_NAME.items()} if torch.cuda.is_available(): device = torch.device("cuda:0") dtype = torch.float16 else: device = torch.device("cpu") dtype = torch.float32 MODEL_NAME = "seamless_expressivity" VOCODER_NAME = "vocoder_pretssel" # used for ASR for toxicity m4t_translator = Translator( model_name_or_card="seamlessM4T_v2_large", vocoder_name_or_card=None, device=device, dtype=dtype, ) unit_tokenizer = load_unity_unit_tokenizer(MODEL_NAME) _gcmvn_mean, _gcmvn_std = load_gcmvn_stats(VOCODER_NAME) gcmvn_mean = torch.tensor(_gcmvn_mean, device=device, dtype=dtype) gcmvn_std = torch.tensor(_gcmvn_std, device=device, dtype=dtype) translator = Translator( MODEL_NAME, vocoder_name_or_card=None, device=device, dtype=dtype, apply_mintox=False, ) text_generation_opts = SequenceGeneratorOptions( beam_size=5, unk_penalty=torch.inf, soft_max_seq_len=(0, 200), step_processor=NGramRepeatBlockProcessor( ngram_size=10, ), ) m4t_text_generation_opts = SequenceGeneratorOptions( beam_size=5, unk_penalty=torch.inf, soft_max_seq_len=(1, 200), step_processor=NGramRepeatBlockProcessor( ngram_size=10, ), ) pretssel_generator = PretsselGenerator( VOCODER_NAME, vocab_info=unit_tokenizer.vocab_info, device=device, dtype=dtype, ) decode_audio = AudioDecoder(dtype=torch.float32, device=device) convert_to_fbank = WaveformToFbankConverter( num_mel_bins=80, waveform_scale=2**15, channel_last=True, standardize=False, device=device, dtype=dtype, ) def normalize_fbank(data: WaveformToFbankOutput) -> WaveformToFbankOutput: fbank = data["fbank"] std, mean = torch.std_mean(fbank, dim=0) data["fbank"] = fbank.subtract(mean).divide(std) data["gcmvn_fbank"] = fbank.subtract(gcmvn_mean).divide(gcmvn_std) return data collate = Collater(pad_value=0, pad_to_multiple=1) AUDIO_SAMPLE_RATE = 16000 MAX_INPUT_AUDIO_LENGTH = 10 # in seconds from pydub import AudioSegment def adjust_audio_duration(input_audio_path, output_audio_path): input_audio = AudioSegment.from_file(input_audio_path) output_audio = AudioSegment.from_file(output_audio_path) input_duration = len(input_audio) output_duration = len(output_audio) # Calcul de la différence de durée duration_diff = input_duration - output_duration # Ajout de silence à la fin si l'audio de sortie est plus court if duration_diff > 0: print("Duration diff : ",duration_diff) silence = AudioSegment.silent(duration=duration_diff) output_audio += silence # Enregistrer l'audio ajusté output_audio.export(output_audio_path, format='wav') return output_audio_path import yt_dlp def dowloadYoutubeAudio(url): print("Téléchargement de l'audio YouTube en cours...") ydl_opts = { 'format': 'm4a/bestaudio/best', 'outtmpl': os.getcwd() + "/audio", # Mise à jour du chemin de sortie 'postprocessors': [{ 'key': 'FFmpegExtractAudio', 'preferredcodec': 'wav', # Utilisation du format WAV }] } with yt_dlp.YoutubeDL(ydl_opts) as ydl: error_code = ydl.download([url]) if error_code == 0: print("Sauvegarde du fichier audio...") print("download_finished : ", os.getcwd() + "/audio.wav") else: print("error : Échec du téléchargement...") return os.getcwd() + "/audio.wav" def split_audio(input_audio_path): print("Start Split Audio") audio = AudioSegment.from_file(input_audio_path) silence_thresh = -20 # Seuil de silence min_silence_len = 300 # Durée minimale de silence en ms chunks = [] current_chunk = AudioSegment.silent(duration=0) for ms in range(0, len(audio), 10): # Incrément de 10 ms segment = audio[ms:ms + 10] current_chunk += segment if len(current_chunk) >= 8000: # Si la durée actuelle dépasse 8 secondes # Vérifier s'il y a un silence if silence.detect_silence(current_chunk[-min_silence_len:], min_silence_len=min_silence_len, silence_thresh=silence_thresh): # Couper au silence print("Silence détecté, découpage du segment") chunks.append(current_chunk) current_chunk = AudioSegment.silent(duration=0) if len(current_chunk) >= 8900: # Si la durée dépasse 9,89 secondes print("Durée maximale atteinte, découpage du segment") chunks.append(current_chunk) current_chunk = AudioSegment.silent(duration=0) # Ajouter le dernier segment s'il n'est pas vide if len(current_chunk) > 0: chunks.append(current_chunk) print('Nombre de segments valides:', len(chunks)) return chunks def remove_prosody_tokens_from_text(text): # filter out prosody tokens, there is only emphasis '*', and pause '=' text = text.replace("*", "").replace("=", "") text = " ".join(text.split()) return text import torchaudio AUDIO_SAMPLE_RATE = 16000 # Taux d'échantillonnage standard def preprocess_audio(input_audio_path: str): print("preprocess_audio start") print("Audio Path :", input_audio_path) audio_segments = split_audio(input_audio_path) temp_folder = os.path.join(os.getcwd(), "path_to_temp_folder") os.makedirs(temp_folder, exist_ok=True) segment_paths = [] for i, segment in enumerate(audio_segments): segment_path = os.path.join(temp_folder, f"segment_{i}.wav") segment_audio = segment.get_array_of_samples() segment_tensor = torch.tensor(segment_audio).unsqueeze(0).float() # Rééchantillonnage segment_tensor = torchaudio.functional.resample(segment_tensor, orig_freq=segment.frame_rate, new_freq=AUDIO_SAMPLE_RATE) torchaudio.save(segment_path, segment_tensor, sample_rate=AUDIO_SAMPLE_RATE) segment_paths.append(segment_path) print("path for :", segment_path) return segment_paths import os import torchaudio # Constante pour le taux d'échantillonnage AUDIO_SAMPLE_RATE = 16000 def preprocess_audio22(input_audio_path: str): print("preprocess_audio start") print("Audio Path :", input_audio_path) # Appeler split_audio et obtenir les segments audio_segments = split_audio(input_audio_path) # Créer un dossier temporaire pour stocker les segments temp_folder = os.path.join(os.getcwd(), "path_to_temp_folder") os.makedirs(temp_folder, exist_ok=True) segment_paths = [] for i, segment in enumerate(audio_segments): # Exporter chaque segment dans un fichier temporaire temp_segment_path = os.path.join(temp_folder, f"temp_segment_{i}.wav") segment.export(temp_segment_path, format="wav") # Charger et traiter le segment audio arr, org_sr = torchaudio.load(temp_segment_path) new_arr = torchaudio.functional.resample(arr, orig_freq=org_sr, new_freq=AUDIO_SAMPLE_RATE) # Enregistrer le segment traité segment_path = os.path.join(temp_folder, f"segment_{i}.wav") torchaudio.save(segment_path, new_arr, sample_rate=AUDIO_SAMPLE_RATE) # Ajouter le chemin du segment traité à la liste segment_paths.append(segment_path) print("Path for :", segment_path) return segment_paths def preprocess_audio222(input_audio_path: str): # Appeler split_audio et obtenir les segments print("preprocess_audio start") print("Audio Path :",input_audio_path) audio_segments = split_audio(input_audio_path) temp_folder = os.getcwd()+"/path_to_temp_folder" os.makedirs(temp_folder, exist_ok=True) segment_paths = [] for i, segment in enumerate(audio_segments): segment_path = os.path.join(temp_folder, f"segment_{i}.wav") segment.export(segment_path, format="wav") segment_paths.append(segment_path) print("path for : ",segment_path) return segment_paths def process_segment(segment_path, source_language_code, target_language_code): # preprocess_audio(segment_path) - cette ligne peut ne pas être nécessaire si le segment est déjà prétraité with pathlib.Path(segment_path).open("rb") as fb: block = MemoryBlock(fb.read()) example = decode_audio(block) example = convert_to_fbank(example) example = normalize_fbank(example) example = collate(example) # Transcription pour mintox source_sentences, _ = m4t_translator.predict( input=example["fbank"], task_str="S2TT", tgt_lang=source_language_code, text_generation_opts=m4t_text_generation_opts, ) source_text = str(source_sentences[0]) prosody_encoder_input = example["gcmvn_fbank"] text_output, unit_output = translator.predict( example["fbank"], "S2ST", tgt_lang=target_language_code, src_lang=source_language_code, text_generation_opts=text_generation_opts, unit_generation_ngram_filtering=False, duration_factor=1.0, prosody_encoder_input=prosody_encoder_input, src_text=source_text, ) speech_output = pretssel_generator.predict( unit_output.units, tgt_lang=target_language_code, prosody_encoder_input=prosody_encoder_input, ) # Chemin pour enregistrer l'audio du segment segment_output_audio_path = os.path.join(os.getcwd(), "result", f"segment_audio_{os.path.basename(segment_path)}") os.makedirs(os.path.dirname(segment_output_audio_path), exist_ok=True) # Enregistrer l'audio du segment torchaudio.save( segment_output_audio_path, speech_output.audio_wavs[0][0].to(torch.float32).cpu(), sample_rate=speech_output.sample_rate, ) segment_output_audio_path = adjust_audio_duration(segment_path, segment_output_audio_path) text_out = remove_prosody_tokens_from_text(str(text_output[0])) print("Audio ici : ",segment_output_audio_path) return segment_output_audio_path, text_out #---------------------------_# from typing import Tuple def run2( input_audio_path: str, source_language: str, target_language: str, ) -> Tuple[str, str]: target_language_code = LANGUAGE_NAME_TO_CODE[target_language] source_language_code = LANGUAGE_NAME_TO_CODE[source_language] preprocess_audio(input_audio_path) with pathlib.Path(input_audio_path).open("rb") as fb: block = MemoryBlock(fb.read()) example = decode_audio(block) example = convert_to_fbank(example) example = normalize_fbank(example) example = collate(example) # get transcription for mintox source_sentences, _ = m4t_translator.predict( input=example["fbank"], task_str="S2TT", # get source text tgt_lang=source_language_code, text_generation_opts=m4t_text_generation_opts, ) source_text = str(source_sentences[0]) prosody_encoder_input = example["gcmvn_fbank"] text_output, unit_output = translator.predict( example["fbank"], "S2ST", tgt_lang=target_language_code, src_lang=source_language_code, text_generation_opts=text_generation_opts, unit_generation_ngram_filtering=False, duration_factor=1.0, prosody_encoder_input=prosody_encoder_input, src_text=source_text, # for mintox check ) speech_output = pretssel_generator.predict( unit_output.units, tgt_lang=target_language_code, prosody_encoder_input=prosody_encoder_input, ) with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f: torchaudio.save( f.name, speech_output.audio_wavs[0][0].to(torch.float32).cpu(), sample_rate=speech_output.sample_rate, ) text_out = remove_prosody_tokens_from_text(str(text_output[0])) return f.name, text_out #---------------------------------------------------------_# #----------------------------------------------------------# #----------------------------------------------__#------ #-----------------------# def run(input_audio_path: str, source_language: str, target_language: str) -> tuple[str, str]: target_language_code = LANGUAGE_NAME_TO_CODE[target_language] source_language_code = LANGUAGE_NAME_TO_CODE[source_language] segment_paths = preprocess_audio22(input_audio_path) print("preprocess_audio end") final_text = "" final_audio = AudioSegment.silent(duration=0) for segment_path in segment_paths: segment_audio_path, segment_text = process_segment(segment_path, source_language_code, target_language_code) final_text += segment_text + " " segment_audio = AudioSegment.from_file(segment_audio_path) final_audio += segment_audio output_audio_path = os.path.join(os.getcwd(), "result", "audio.wav") os.makedirs(os.path.dirname(output_audio_path), exist_ok=True) final_audio.export(output_audio_path, format="wav") text_out = remove_prosody_tokens_from_text(final_text.strip()) return output_audio_path, text_out TARGET_LANGUAGE_NAMES = [ "English", "French", "German", "Spanish", ] from flask import Flask, request, jsonify import torch import torchaudio app = Flask(__name__) # Fonction run adaptée pour Flask @app.route('/translate', methods=['POST']) def translate(): # Récupérer les données de la requête data = request.json input_audio_path = data['input_audio_path'] source_language = data['source_language'] target_language = data['target_language'] # Exécution de la fonction de traduction output_audio_path, output_text = run(input_audio_path, source_language, target_language) # Retourner la réponse return jsonify({ 'output_audio_path': output_audio_path, 'output_text': output_text }) import os url = "https://youtu.be/qb_tHWGJOp8?si=10qB2JApy0q3XY76" input_audio_path = dowloadYoutubeAudio(url) #input_audio_path = os.getcwd()+"/au1min_Vocals_finale.wav" source_language = "French" target_language = "English" print("Audio à traiter : ",input_audio_path) output_audio_path, output_text = run(input_audio_path, source_language, target_language) print("output_audio_path : ",output_audio_path)