SteerTTS: a 23M open TTS you steer with numbers, on CPU
Twelve z-score dials and 2,825 voices let an LLM shape pitch and pace live, real-time on CPU — weights are non-commercial. Plus: Anthropic buys Opaque.
Copy markdown
A tiny voice model that takes direction
SteerTTS v1 is 23.1M params — a 9.59M acoustic model plus a Vocos vocoder — and streams in real time on CPU: RTFx ~4.9 on 4 threads, first audio in ~0.5s, no GPU required. English-only, 24kHz output, with word error rate around 1.8%.
Twelve dials, measured in standard deviations
You set 12 acoustic controls — rate, loudness, median pitch, pitch range, spectral tilt and more — as z-scores from -5 to +5 against a voice's neutral baseline; omit one and the model fills it in. Six track reliably (Spearman rho >= 0.85). It ships 2,825 built-in voices, inline markup like <slow> and [pause 300ms], and sentence-level streaming.
Why it matters: a voice your agent can modulate mid-sentence
Because every knob is a number, your LLM can emit the prosody itself — slow down for a warning, lift the pitch for a question — with no separate voice API and no GPU bill. A Mamba-3 hybrid variant (steer-tts-m3hyb) adds per-word pitch, energy and duration edits plus 33 style tags.
Read the license before you ship
Weights are CC BY-NC 4.0 — non-commercial only (the package code is MIT) — because training used non-commercial data. Cloning works from a short clip, but the card warns "do not expect a convincing likeness" (33.1 avg speaker similarity), and emotion tags mostly behave as style switches.
Also today: Anthropic acqui-hires Opaque Systems
Anthropic licensed confidential-AI startup Opaque's tech and brought on its CEO, CTO and engineers; Opaque winds down. The deal adds hardware-attested confidential computing and "TRACE" agent-governance tooling — a marker of where controls for deploying agents in regulated settings are heading.