| 1 | import os |
| 2 | import sys |
| 3 | |
| 4 | |
| 5 | sys.path.append(os.getcwd()) |
| 6 | |
| 7 | import json |
| 8 | from importlib.resources import files |
| 9 | from pathlib import Path |
| 10 | |
| 11 | import soundfile as sf |
| 12 | from datasets.arrow_writer import ArrowWriter |
| 13 | from tqdm import tqdm |
| 14 | |
| 15 | |
| 16 | def main(): |
| 17 | result = [] |
| 18 | duration_list = [] |
| 19 | text_vocab_set = set() |
| 20 | |
| 21 | with open(meta_info, "r") as f: |
| 22 | lines = f.readlines() |
| 23 | for line in tqdm(lines): |
| 24 | uttr, text, norm_text = line.split("|") |
| 25 | norm_text = norm_text.strip() |
| 26 | wav_path = Path(dataset_dir) / "wavs" / f"{uttr}.wav" |
| 27 | duration = sf.info(wav_path).duration |
| 28 | if duration < 0.4 or duration > 30: |
| 29 | continue |
| 30 | result.append({"audio_path": str(wav_path), "text": norm_text, "duration": duration}) |
| 31 | duration_list.append(duration) |
| 32 | text_vocab_set.update(list(norm_text)) |
| 33 | |
| 34 | # save preprocessed dataset to disk |
| 35 | if not os.path.exists(f"{save_dir}"): |
| 36 | os.makedirs(f"{save_dir}") |
| 37 | print(f"\nSaving to {save_dir} ...") |
| 38 | |
| 39 | with ArrowWriter(path=f"{save_dir}/raw.arrow") as writer: |
| 40 | for line in tqdm(result, desc="Writing to raw.arrow ..."): |
| 41 | writer.write(line) |
| 42 | writer.finalize() |
| 43 | |
| 44 | # dup a json separately saving duration in case for DynamicBatchSampler ease |
| 45 | with open(f"{save_dir}/duration.json", "w", encoding="utf-8") as f: |
| 46 | json.dump({"duration": duration_list}, f, ensure_ascii=False) |
| 47 | |
| 48 | # vocab map, i.e. tokenizer |
| 49 | # add alphabets and symbols (optional, if plan to ft on de/fr etc.) |
| 50 | with open(f"{save_dir}/vocab.txt", "w") as f: |
| 51 | for vocab in sorted(text_vocab_set): |
| 52 | f.write(vocab + "\n") |
| 53 | |
| 54 | print(f"\nFor {dataset_name}, sample count: {len(result)}") |
| 55 | print(f"For {dataset_name}, vocab size is: {len(text_vocab_set)}") |
| 56 | print(f"For {dataset_name}, total {sum(duration_list) / 3600:.2f} hours") |
| 57 | |
| 58 | |
| 59 | if __name__ == "__main__": |
| 60 | tokenizer = "char" # "pinyin" | "char" |
| 61 | |
| 62 | dataset_dir = "<SOME_PATH>/LJSpeech-1.1" |
| 63 | dataset_name = f"LJSpeech_{tokenizer}" |
| 64 | meta_info = os.path.join(dataset_dir, "metadata.csv") |
| 65 | save_dir = str(files("f5_tts").joinpath("../../")) + f"/data/{dataset_name}" |
| 66 | print(f"\nPrepare for {dataset_name}, will save to {save_dir}\n") |
| 67 | |
| 68 | main() |
| 69 |