SteerTTS: 23M-parameter English TTS model with 12 numeric controls and real-time CPU inference

·2026.10.19 14:36

Key point

The model achieves ~1.8% average WER on Open TTS Leaderboard English sets and supports 51 style tags plus zero-shot cloning.

23M-parameter English TTS model combining a 9.6M-param acoustic model (attention + bidirectional Mamba-3, FastPitch-style) and a Vocos vocoder.
12 controls (rate, loudness, pitch, etc.) operate as z-scores relative to the voice's neutral speech, with 2,825 built-in anonymized voices.
Real-time performance with RTFx ~60–80 on RTX 5070 and ~4–8 on CPU; first streamed audio in ~0.5 s on CPU.
~1.8% average WER on Open TTS Leaderboard English sets (local run, not official submission).
Limitations include weak pause/voicing controls, low cloning similarity, and English-only support.
Weights licensed under CC BY-NC 4.0 (non-commercial) and code under MIT; available on Hugging Face.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.