import torch import torchaudio import time import os import numpy as np import json from datasets import load_dataset, Audio from snac import SNAC from torch.nn import functional as F from tqdm import tqdm import wandb # Constants SNAC_SAMPLE_RATE = 24000 OUTPUT_DIR = "processed_common_voice" BATCH_SIZE = 1000 # Ensure CUDA is available device = torch.device("cuda" if torch.cuda.is_available() else "cpu") def load_snac_model(sample_rate): if sample_rate == 24000: model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval().to(device) else: raise ValueError("Unsupported sample rate. Please use 24000.") return model snac_model = load_snac_model(SNAC_SAMPLE_RATE) def chunk_and_pad_audio(audio, chunk_size): length = audio.shape[-1] padded_length = ((length + chunk_size - 1) // chunk_size) * chunk_size padded_audio = F.pad(audio, (0, padded_length - length), mode="constant", value=0) batched_audio = padded_audio.unfold(-1, size=chunk_size, step=chunk_size) return batched_audio def generate_snac_encoding(audio): waveform = torch.tensor(audio["array"]).float().to(device) if audio["sampling_rate"] != SNAC_SAMPLE_RATE: resampler = torchaudio.transforms.Resample( orig_freq=audio["sampling_rate"], new_freq=SNAC_SAMPLE_RATE ) waveform = resampler(waveform) if waveform.dim() == 2: waveform = waveform.mean(dim=0, keepdim=True) elif waveform.dim() == 1: waveform = waveform.unsqueeze(0) num_second = 1 chunk_size_initial = num_second * SNAC_SAMPLE_RATE lcm = np.lcm.reduce([snac_model.vq_strides[0], snac_model.attn_window_size or 1]) pad_to = snac_model.hop_length * lcm chunk_size = int(np.ceil(chunk_size_initial / pad_to) * pad_to) audio = chunk_and_pad_audio(waveform, chunk_size) audio = audio.permute(1, 0, 2) codes_list = [] with torch.no_grad(): for chunk in audio: codes = snac_model.encode(chunk.unsqueeze(0)) codes = [c.cpu() for c in codes] codes_list.append(codes) codes_list = [torch.cat(codes_list, dim=0) for codes_list in zip(*codes_list)] codes_list = [code.reshape(-1).cpu().tolist() for code in codes_list] string_codes = " ".join(map(str, codes_list[0])) return string_codes def process_audio(item): start_time = time.time() try: snac_tokens = generate_snac_encoding(item["audio"]) if not snac_tokens: raise ValueError("Generated SNAC tokens are empty") except Exception as e: return None processing_time = time.time() - start_time return { "path": item["path"], "sentence": item["sentence"], "age": item["age"], "gender": item["gender"], "accent": item["accent"], "locale": item["locale"], "snac": snac_tokens, "processing_time": processing_time, "audio_duration": len(item["audio"]["array"]) / item["audio"]["sampling_rate"], } def save_to_jsonl(data, file_path): # Open the file in append mode to add new data to the existing language-specific JSONL file with open(file_path, "a") as f: for item in data: json.dump(item, f) f.write("\n") def process_language(language): # Ensure output directory exists language_dir = os.path.join(OUTPUT_DIR, language) os.makedirs(language_dir, exist_ok=True) jsonl_path = os.path.join(language_dir, f"{language}_processed.jsonl") # Read existing data existing_data = set() if os.path.exists(jsonl_path): with open(jsonl_path, "r") as f: existing_data = set(f.readlines()) # Load the Common Voice dataset for this language dataset = load_dataset( "mozilla-foundation/common_voice_16_1", language, split="train", streaming=True ) # Cast the dataset to include audio dataset = dataset.cast_column("audio", Audio(sampling_rate=SNAC_SAMPLE_RATE)) processed_data = [] total_processed = 0 report_counter = 0 for item in tqdm(dataset, desc=f"Processing {language}"): result = process_audio(item) if result: json_line = json.dumps(result) + "\n" if json_line not in existing_data: processed_data.append(result) existing_data.add(json_line) total_processed += 1 report_counter += 1 if report_counter % 1000 == 0: # Report to wandb every 1000 rows wandb.log( { "language": language, "average_processing_time": np.mean( [item["processing_time"] for item in processed_data] ), "average_audio_duration": np.mean( [item["audio_duration"] for item in processed_data] ), "average_snac_token_count": np.mean( [len(item["snac"].split()) for item in processed_data] ), } ) report_counter = 0 # Reset the counter # Save every BATCH_SIZE items if len(processed_data) >= BATCH_SIZE: save_to_jsonl(processed_data, jsonl_path) processed_data = [] # Clear the list after saving # Save any remaining processed data if processed_data: save_to_jsonl(processed_data, jsonl_path) return total_processed def main(): # Initialize wandb wandb.init(project="common-voice-processing", job_type="data-processing") # List of languages to process, starting with English languages = ['ckb', 'cnh', 'cs', 'cv', 'cy', 'da', 'de'] # languages = ['dv', 'dyu', 'el', 'en', 'eo', 'es', 'et'] # languages = ['eu', 'fa', 'fi', 'fr', 'fy-NL', 'ga-IE', 'gl'] # languages = ['gn', 'ha', 'he', 'hi', 'hsb', 'hu', 'hy-AM'] # languages = ['ia', 'id', 'ig', 'is', 'it', 'ja', 'ka'] # languages = ['kab', 'kk', 'kmr', 'ko', 'ky', 'lg', 'lij'] # languages = ['lo', 'lt', 'ltg', 'lv', 'mdf', 'mhr', 'mk'] # languages = ['ml', 'mn', 'mr', 'mrj', 'mt', 'myv', 'nan-tw'] # languages = ['ne-NP', 'nhi', 'nl', 'nn-NO', 'oc', 'or', 'os'] # languages = ['pa-IN', 'pl', 'ps', 'pt', 'quy', 'rm-sursilv', 'rm-vallader'] # languages = ['ro', 'ru', 'rw', 'sah', 'sat', 'sc', 'sk'] # languages = ['skr', 'sl', 'sq', 'sr', 'sv-SE', 'sw', 'ta'] # languages = ['te', 'th', 'ti', 'tig', 'tk', 'tok', 'tr'] # languages = ['tt', 'tw', 'ug', 'uk', 'ur', 'uz', 'vi', 'vot', 'yi', 'yo', 'yue', 'zgh', 'zh-CN', 'zh-HK', 'zh-TW'] total_processed_all_languages = 0 # Process each language for language in languages: total_processed = process_language(language) total_processed_all_languages += total_processed print( f"\nCompleted processing all languages. Total files processed across all languages: {total_processed_all_languages}" ) wandb.finish() if __name__ == "__main__": main()