TL;DR: WhissleAI released an 11.84 MB int4 edge variant of SteerTTS that keeps WER unchanged while dropping UTMOS by only ~0.1, making steerable TTS practical for on-device apps.
Summary: WhissleAI released SteerTTS Edge, a compact English text-to-speech model whose full app payload—weights, voices, text front end, and Python runtime—totals 11.84 MB. The acoustic model uses 9.25M parameters with attention plus bidirectional Mamba-3 and is quantization-aware trained to int4, while a 6.2M-parameter Vocos-style vocoder is distilled and also int4 QAT; the text front end is pure NumPy with no NLTK, g2p_en, torchaudio, or num2words at runtime. On the Open TTS Leaderboard protocol run locally, Seed-TTS WER stays at 0.96% and UTMOS falls from 4.20 to 4.11, with all 12 steering controls retaining at least 95% of their fp32 effect; it runs at RTF 0.053 on CPU with four threads and offers 50 style tags plus inline markup, but is English-only, limited to 200 voices, and does not support voice cloning.
Why it matters: Tiny, steerable, CPU-friendly TTS enables offline voice interfaces and privacy-preserving on-device apps without cloud latency. Try the edge model for latency-sensitive deployments, but account for its English-only coverage and lack of voice cloning.
Source: reddit