You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

このデータセットは moe-speech-plus(情報解析目的限定・再配布禁止)から派生した特徴量 (テキスト・離散音声トークン・話者埋め込み)です。研究・情報解析目的でのみ利用し、 再配布しないことに同意してください。This dataset is derived features of moe-speech-plus (research/text-and-data-mining use only, no redistribution). By requesting access you agree to research-only use and no redistribution.

Log in or Sign Up to review the conditions and access this dataset content.

moe-speech-plus CosyVoice3 waveform-free parquet(llm 学習用特徴量)

ayousanz/moe-speech-plus(日本語キャラクター演技音声)を CosyVoice3 の llm 学習用に前処理した波形フリー parquet です。音声バイトは含みません。

  • 289,893 発話(≒470h 相当、speechMOS≥1.5 + cross-CER≤0.2 フィルタ済み)、train 290 / dev 8 シャード
  • 列: utt / text(漢字混じり原文)/ instruct / utt_embedding・spk_embedding(campplus 192次元)/ speech_token(speech_tokenizer v3、25Hz、FSQ 6561)
  • これがあれば CosyVoice3 系の llm 学習(AR FT・DELTA-TTS delta 変換とも)で 元データ 326GB のダウンロードと約2.5時間の前処理を丸ごとスキップできます

使い方

tar xzf delta_data_artifacts.tgz -C examples/moe_speech/cosyvoice3/
# -> data/train/parquet data/dev/parquet data/train.data.list data/dev.data.list が復元される
# (data.list は相対パスなのでリポジトリ配置が同じならそのまま学習に使える)

学習レシピ: ayutaz/CosyVoice feature/delta-tts-reproduction (examples/moe_speech/cosyvoice3/、AR FT は run.sh、delta 変換は scripts/prod/vast_delta_run.sh / vast_delta_refine.sh)。tgz には学習時の tensorboard ログと yaml も同梱しています。

ライセンス

元データ moe-speech-plus の条件(情報解析目的限定・再配布禁止)を継承します。

Downloads last month
7