[leehack]

Summary

• Speak with Chatterbox Turbo, Chatterbox Multilingual and Pocket TTS (ChatterboxTtsAdapter, PocketTtsAdapter) through the native model-type opt-in, using llamadart-native v0.9.0 types 4 and 5. Refs Expose Chatterbox and Pocket TTS voices from native v0.9.0 #1783.
• Predefined voices come from the runtime listing added in Speak with LFM3-Audio and Soprano through the native model-type opt-in #1791 (merged): Chatterbox default, the 25 Pocket TTS voices. Speaker-reference cloning works for Chatterbox (and for Pocket GGUFs that ship a voice encoder), and Multilingual takes language.
• Old runtimes refuse before loading any file: v0.8.0 and older throw LlamaUnsupportedException naming llama_dart_tts_supported_model_types and llamadart-native v0.9.0.
• Targets main now that Speak with LFM3-Audio and Soprano through the native model-type opt-in #1791 (Expose LFM3 and Soprano speech output from native v0.8.0 #1615) is merged; main pins llamadart-native v0.9.0. The diff against main is only this PR's families plus the probe tests below.
• Adds the two test gaps from the Speak with LFM3-Audio and Soprano through the native model-type opt-in #1791 audit: the llama.cpp worker and backend client answer the pre-load model-support probe (worker_test.dart, llama_cpp_backend_test.dart), and the Web and WebGPU backends refuse non-Qwen3 models before loading and accept Qwen3-TTS (Chrome, web_backend_test.dart, webgpu_backend_test.dart). Each fails when its code path is nulled.

Production-readiness scope

• User-facing scope: Chatterbox Turbo and Multilingual speech, voice default or speaker-reference cloning, Multilingual language; Pocket TTS speech with 25 predefined voices. The example CLI gains the two adapters plus --speaker and --sequences.
• Supported platforms/paths: Native llama.cpp runtimes that report types 4/5 in llama_dart_tts_supported_model_types (llamadart-native v0.9.0 or later).
• Unsupported or intentionally unavailable paths:
  • v0.8.0 and older, or Apple companion locks on them: load throws LlamaUnsupportedException naming the symbol and v0.9.0.
  • Chatterbox Multilingual in a context with fewer than two sequences: capabilities report unsupported, asking for ModelParams.maxParallelSequences of at least 2 (the classifier-free-guidance twin sequence).
  • Voice plus speaker reference: LlamaTextToSpeechException before synthesis. Turbo and Pocket TTS: language throws LlamaUnsupportedException. Pocket without a voice encoder: speaker reference throws LlamaUnsupportedException.
  • Web/WebGPU: load refuses these families (Qwen3-TTS only).
  • Adreno 750 Vulkan stays off through the existing Android driver guard.
• Out of scope / follow-ups: None.

Completeness checklist

• Declared scope is fully implemented, or explicitly reduced above.
• Unsupported platform/option combinations fail loudly with actionable diagnostics or are clearly documented as unavailable.
• Public API docs, README/website docs, examples, support matrices, and changelog entries are updated where relevant.
• Regression coverage covers the original issue plus key negative/version-skew paths where relevant.
• Public capability changes update the validation coverage registry with a pack or a reasoned, issue-linked exemption; inventory checks pass (or N/A with scope). No new capability fields; the families are covered by the text-to-speech-models-smoke exemption reason from PR1.
• Security/privacy review completed: no secrets, bearer tokens, signed URLs, or raw secret-bearing paths leak through logs, cache keys, metadata, errors, or snapshots.
• Follow-up work that is useful but not required for this PR is tracked in GitHub Issues, or explicitly marked "None" above.

PR type guidance

• Feature PR: API, docs, example, support matrix, unsupported paths and positive/negative tests are included above.
• Bugfix PR: N/A, feature.
• Docs-only PR: N/A, runtime change.

Test Plan

• dart run tool/prepare_workspace.dart
• dart format --output=none --set-exit-if-changed .
• dart analyze
• dart test -p vm -j 1 --exclude-tags local-only: full suite 11518 pass, 129 skip, 6 fail (all windows_registry_qualification_test.dart: no pwsh on this host) at 672b0de; at this head the TTS, worker, backend, entrypoint, structure and tooling suites pass (692 tests), and PR CI runs the full suite.
• dart test -p chrome --exclude-tags local-only: full suite 4626 pass, 1 skip before the retarget; at this head the Web, WebGPU and speech suites pass (391 tests), and PR CI runs the full suite.
• coverage-lib matrix row: not run locally; main enforces it after merge.
• Other targeted/local-only validation: text-to-speech-models-smoke, llama-cpp-exit-teardown (quit-speaking), version skew on v0.8.0-3.

Matrix Evidence

Matrix row	Scope covered	Platform / model / backend	Result	Evidence / notes
text-to-speech-models-smoke (published v0.9.0, this head)	Chatterbox Turbo, Multilingual, Pocket TTS after merging main with #1791	macOS arm64 M4 Max, published llamadart-native v0.9.0 (hook output sha256 9d9cdbc6…), no local override, CPU, 2/2 threads, isolated TMPDIR	Pass (20 pass, 25 skip: LFM3/Soprano unset plus inapplicable cases)	Transcripts exact for Turbo default and cloned, Multilingual default and cloned, Pocket default and vera; Multilingual fr gives "Bonjour, comment allez-vous aujourd'hui?".
text-to-speech-models-smoke (published v0.9.0, all families)	All five families after merging main (#1784 landed, pin v0.9.0); no local native override	macOS arm64 M4 Max, published llamadart-native v0.9.0 bundle (hook output sha256 9d9cdbc6…), CPU, 2/2 threads, isolated TMPDIR	Pass (29 pass, 16 inapplicable skips)	ASR transcripts exact for LFM3 default and male, Soprano, Chatterbox Turbo default and cloned, Multilingual default and cloned, Pocket default and vera; Multilingual fr gives "Bonjour, comment allez-vous aujourd'hui?". Run at 672b0de, before the retarget.
text-to-speech-models-smoke (pre-release native)	Chatterbox Turbo, Multilingual, Pocket TTS: capabilities, voices, cloning, language, refusals, seed repeat after cancel, Qwen3-ASR round trip	macOS arm64 M4 Max, local llamadart-native main bee747a5, CPU, 2/2 threads; ggml-org/chatterbox-turbo-GGUF Q8_0, ggml-org/chatterbox-GGUF Q8_0 (F16 mmprojs), macolomb/pocket-tts-without-voice-cloning english; ASR ggml-org/Qwen3-ASR-0.6B-GGUF Q8_0	Pass (20 pass, 25 skip: LFM3/Soprano unset in that run plus inapplicable cases)	Transcripts exact for Turbo default and cloned, Multilingual default and cloned, Pocket default and vera; Multilingual fr gives "Bonjour, comment allez-vous aujourd'hui?". Pocket lists 25 voices; no voice equals the first voice; Multilingual refuses a one-sequence context.
version skew: v0.8.0-3	pre-load refusal	pinned v0.8.0-3, example CLI	Pass	Chatterbox and Pocket TTS throw LlamaUnsupportedException naming v0.9.0 before loading; LFM3/Soprano still speak.
llama-cpp-exit-teardown (quit-speaking)	exit() mid-synthesis	macOS arm64, local llamadart-native main bee747a5 (pre-release), Metal	Pass	{"quit-speaking":[0,0,0]} for Chatterbox Multilingual and Pocket TTS.
probe mutation checks	worker, backend client, Web and WebGPU answers to the pre-load probe	VM and Chrome	Pass	Nulling the worker's reason, the backend client's return, either Web backend's override, or refusing Qwen3 in the shared Web function each fails the matching new test.
model-free integration	type mask 0x3e, Multilingual sequence refusal, Turbo/Pocket capability flags, version messages	VM, mirror-injected fake _TtsApi	Pass	test/integration/backends/llama_cpp_tts_model_types_test.dart, text_to_speech_model_test.dart.
llama-cpp-host-shutdown (Linux)	N/A	-	Not run	No Linux host.
chat-app-macos-quit	N/A	-	Not run	Chat app TTS is Qwen3-only and unchanged.
coverage-lib	N/A	-	Not run	Enforced on main after merge.
Windows / Linux / Android / iOS real-model	N/A	-	Not run	No such hosts or devices here.

Review Notes

• Independent review status: APPROVE from a fresh independent auditor at exact head 99891bf160a1f2dbe7ab9de00812d128942495b6 against base ce8ddf674200297e5883d74e8888552c4af612b1; 0 P1/P2. high_risk_readiness.dart exits 2 (unverifiedPrerequisites).
• CI status / head SHA: 99891bf160 (merge of main after Speak with LFM3-Audio and Soprano through the native model-type opt-in #1791 landed as ce8ddf6), PR CI https://github.com/leehack/llamadart/actions/runs/42484282344 green: 34 checks pass, 9 skipped. The probe-test commit b161d90 also passed a dispatched run: https://github.com/leehack/llamadart/actions/runs/42483386669.

High-risk regression review

• Classification: high-risk (backendRuntime, regressionPolicy).
• Implementation task: this PR's author session.
• Independent blocking QA task: done, APPROVE; VM 410 targeted and 3392 broad (1 unrelated LiteRT flake under -j 4), Chrome 383, coverage and skills 48.
• Exact head / current base: 99891bf160a1f2dbe7ab9de00812d128942495b6 / ce8ddf674200297e5883d74e8888552c4af612b1.
• Production-branch deletion, bypass, or miswire proof: auditor killed 17 of 17 mutants (multilingual gate, version map, adapter supportsModel, probe routing through worker, backend, _runtimeRefusal, Web, WebAutoBackend, WebGpu); real v0.8.0 refuses Chatterbox and Pocket before load. The integration suite fails if the type mask drops types 4/5 or the Multilingual sequence check is bypassed.
• Affected-family real-model/artifact evidence: Chatterbox Turbo, Chatterbox Multilingual, Pocket TTS, see Matrix Evidence.
• Explicit unavailable-family or other N/A evidence: Pocket TTS cloning with a voice-encoder GGUF is not exercised: the GGUF used ships no encoder, and the E2E asserts its speaker-reference refusal. Dart reads the native SPEAKER_REFERENCE bit, so no family-specific code path is untested.
• Known PR-caused P1 regressions: 0
• Unresolved review threads: 0
• Audit P3s (not blocking): the exit-probe fixture now loads every adapter with maxParallelSequences: 2; Soprano was not run on real weights.

[github-actions]

Chat app preview removed for leehack/llamadart-chat-pr-1792.