Skip to content

Latest commit

 

History

History
40 lines (29 loc) · 1.31 KB

File metadata and controls

40 lines (29 loc) · 1.31 KB
domain contrib
title gpt sovits name2text arpabet
verification metadata-normalized
{"title" GPT-SoVITS 训练:2-name2text 格式必须用 ARPABET 音素而非中文原文", "domain": "tts", "tags": "", "source": "hanged-man", "status": "published", "created": "2026-04-06", "confidence": "0.9", "scope": "narrow", "domain_expert": "hanged-man", "verified_date": "2026-04-06"}
created 2026-07-06
source unknown

问题

训练时数据加载器逐字查 phoneme 词典,全部 KeyError。

根因

2-name2text.txt 第二列误写为中文原文,正确应为 ARPABET 音素符号(空格分隔)。

正确格式

basename	{w o2 h en3 AA ai4 ...}	{type}	{language}

注意:

  • 中文→ARPABET:用 g2p(text_normalize(text))
  • 必须先 text_normalize 再 g2p,中文标点( )需先规范化为 ASCII 标点
  • 音频文件必须加 .wav 扩展名,无扩展名的 WAV ffmpeg 无法识别

Verification

  1. Follow the solution steps in order
  2. Run any relevant commands or tests to confirm the fix
  3. Verify the symptom no longer occurs
  4. Check related logs or outputs for expected behavior

教训

音素训练数据格式必须严格按文档,词典只认音标不认文字。