Repository navigation
feat(tts): register voxcpm2, zonos2, chatterbox and higgs engines for evaluation - #260
Merged
Merged
Conversation
…tion Adds three EngineSpecs and adapters for the 2026-10-11 engine evaluation and a per-engine sample_rate (48k / 44.1k / 24k) that _render_take writes the take at, dying if the loaded model reports another. qwen3 and breeze are unchanged. No show selects a new engine. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Clone + events: (laugh)/(sigh) become Higgs' inline sfx tags at the adapter seam. EngineSpec gains model_type, passed to load_model only when set, because mlx-audio 0.5.8 cannot resolve Higgs v3's config model_type from the repo id. qwen3's load call is unchanged. No show selects the engine. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Owner
Author
|
Internal change — no release note. The new engines are registered for evaluation only and no show selects them, so listeners and existing manifests see no difference. Generated by Claude Code |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Requested by Schmug
Summary
Registers three open-weight TTS engines in
render.ENGINESso the eval bench can measure them against theqwen3control:voxcpm2(VoxCPM2 8-bit, Apache 2.0),zonos2(Zyphra ZONOS2, Apache 2.0) andchatterbox(Chatterbox Multilingual v3, MIT). Each has an adapter in_ENGINE_GENERATORS. These three were picked from a survey of models released or updated in the last four months that clone fromref_audio, allow a public podcast, and run in mlx-audio.EngineSpecgainssample_rate(default 24000), because the new models output 48 kHz, 44.1 kHz and 24 kHz._render_takewrites each take at the engine's declared rate. It dies if the loaded model reports a different rate, since a wrong rate would pitch-shift every take silently. Qwen3 and Breeze both report 24000, so their writes are unchanged and the qwen3 kwargs are untouched.A follow-up commit adds
higgs(Higgs Audio v3,bosonai/higgs-tts-3-4b) at Schmug's request. It can clone and perform events:(laugh)and(sigh)are rewritten to Higgs' inline sfx tags in the adapter. It also adds anEngineSpec.model_typefield, passed toload_modelonly when set, because mlx-audio 0.5.8 cannot resolve Higgs' architecture from the repo id. Its weights are non-commercial except under Boson's creator-use grant.No show's
tts_enginechanges. The bench results and the recommendation go in the thread, not this PR.Related issue
Refs #200 (eval bench). There is no tracking issue for this evaluation.
Acceptance criteria
EngineSpecfollowing the 2026-09-04 registry designTesting
ruff check .andruff format --check .are green (ruff 0.14.10)pytest: 1787 passed, 1 skippedstatus: dry-run,tts_engine: qwen3, takes mono 44.1k, loudnorm -24.04 LUFSsample_rateits spec declares (qwen3 24000, breeze 24000, voxcpm2 48000, zonos2 44100, chatterbox 24000)Risk / blast radius
Every take now passes through
spec.sample_rateinstead ofSAMPLE_RATE. That value is 24000 for qwen3, so the write is the same as before. The new die-on-mismatch check only fires when a model reports a rate other than the one its spec declares. The real qwen3 and breeze models both report 24000.Docs kept in sync
skills/daily-podcast/SKILL.md(engine table, schema comment), the Surface Tension TTS rules table, and CLAUDE.md's engine sentence🤖 Generated with Claude Code