Can one model
normalize and speak?
From complex written text to spoken language.
Listen to end-to-end normalization, coverage-aware learning,
and explicit reading-position supervision.
**PRO** · 39元/月
$4 \times 12 = 48$
18℃–20℃
Aemeath
Anchored Talker + MTP adaptation
Everbright
Coverage-aware policy learning
Polestar
Explicit reading-position supervision
The same text. Different readings.
12 model / variant configurations
RAW INPUT
Shared by all modelsACCEPTED READING
Human referenceListen & compare
CER compares this clip’s ASR transcript with the accepted reading. Lower is better.
How are the systems compared?
Each example uses the same raw input across systems. Text frontends are labeled: IndexTTS2 with TN on/off, CosyVoice3 and VoxCPM2 with TN off, FireRedTTS3 with WeText, and XTTS-v2 with its native frontend. Systems retain their native voices and decoding configurations.
What do the reference and CER mean?
The reference provides human-annotated accepted readings. Forms inside [A|B|C] are alternatives; [X] is optional. CER is calculated from Qwen3-ASR transcripts using case-insensitive normalization and accepted readings. Use the audio players to compare pronunciation, reading completeness and voice quality.
How can I explore the audio?
All 180 audio clips are original model outputs. Use Quick A/B to compare two systems on the same input, or download individual WAV files. Sample texts, accepted readings, ASR transcripts and audio paths are available in the listening manifest.