Fish Audio Drama 3: open-weight 35B TTS you direct in plain text
Downloadable weights, 16 speaker tags per script, 83 languages, and emotion that shifts mid-line. Plus plannotator, a review gate for Claude Code and Codex.
Copy markdown
Drama 3 drops with open weights
Fish Audio open-weighted Drama 3, a 35B mixture-of-experts voice model (3B active per token) spanning 83 languages. Pull it with `hf download fishaudio/drama-3` and serve it through SGLang-Omni or vLLM-Omni, which expose an OpenAI-style /v1/audio/speech endpoint — a near drop-in for existing TTS code. The license is a revenue-capped community license, so hobby and small-team use is free.
Direct the read like a voice actor
Steer delivery with plain-text notes mid-line, shift emotion inside a single sentence, and drop inline event tags like [laughs]. One script can carry up to 16 speaker tags, clone a voice from a short reference clip, and regenerate a single wrong word with audio inpainting instead of re-recording the whole take. Multi-character audiobooks, podcasts, and game dialogue become a weekend build.
plannotator: gate your agent's plan before it runs
plannotator is a local, browser-based review surface for coding agents: annotate the plan it proposes or the diff it produced, then fire feedback back with one click. It hooks into Claude Code, Codex, Gemini CLI (0.36+), Copilot CLI, Amp, and OpenCode, and reviews local changes or remote PRs across Git, GitHub, and GitLab. It is Apache-2.0/MIT with a one-line installer.