SteerTTS-edge: an 11.84MB steerable TTS with unchanged WER running 19x realtime on CPU
Working_Hat5120 · reddit · 2026-10-23
WhissleAI released SteerTTS-edge, shrinking everything an app must ship (weights, voices, text frontend, Python runtime) to 11.84MB, open on Hugging Face.
How: the 9.25M-param acoustic model (attention + bidirectional Mamba-3) and a 6.2M-param Vocos-style vocoder are both int4 quantization-aware trained; the text frontend is pure numpy with no NLTK/g2pen/torchaudio at runtime.
Quality: Seed-TTS WER stays at 0.96%, UTMOS drops 0.1 (mostly from the int4 vocoder). All 12 numeric controls (rate, loudness, pitch, pauses, breathiness, etc.) retain at least 95% of their fp32 effect; 50 style tags and inline markup like <slow> and <whisper> survive too.
Speed: CPU-only RTF 0.053 (19x realtime), 120ms to first audio, sentence-level streaming.
Limits: English only, 200 voices, no voice cloning; weights dequantize to fp32 at load so the win is size, not speed. Feedback wanted on vocoder quality and ONNX/mobile/WASM targets.
More from Multimodal
- Ant's Ming-3.0-tts hits top-tier TTS quality with just 0.5B params, MIT licensed — ArtificialAnlys · 2026-10-23
- Suno Highlights Five Niche Features Most Users Don't Know About — suno · 2026-10-23
- Meshy offers Hermes users a month of Pro with API access for $1 — Teknium · 2026-10-23
- Customuse launches Smart Mesh AI with geometry-aware adaptive character clothing tool — SimplyAnnisa · 2026-10-23
- AI-generated movie poster 'The Chase' crams a wolf, a plane and muscle cars into one chase — Aggravating-Sir-115 · 2026-10-23
- ElevenLabs' Chimera pairs Eleven v4 voice with Seedance video to make character images speak — AIwithGhotai · 2026-10-23