Update CrispASR to v0.8.25 and add GigaAM-v3 (STT) + OmniVoice TTS (CrispASR)

[niksedk]

Three commits: the runtime bump, then the two new engines it makes available.

Update CrispASR to v0.8.25

Bumps the pinned runtime across all 11 platform download URLs and adds the new archive + executable SHA-256 hashes, all computed from freshly downloaded release assets (sizes match the release API exactly).

Verification:

• macOS binary probe-tested: reports 0.8.25, git sha 6e5bf7ff, backends cpu,metal,blas, and transcribes correctly via the sensevoice backend on Metal.
• Diffed the full --help flag set against the installed v0.8.24 binary: identical, so nothing SE passes was removed or renamed. Upstream confirms it is drop-in.
• Download-size figures refreshed from the v0.8.25 assets.

Notable upstream fixes for SE users: the pyannote powerset decode table had two entries transposed, so every frame with a third speaker talking alone was credited to the wrong two (15 DER points on VoxConverse dev); --diarize-max-speakers was picking the speaker count instead of bounding it (15.74% → 7.81% DER with TitaNet); pyannote segmentation is now chunked and parallel (~50s → ~18s on a 2888s file at -t 8); sherpa diarization no longer hangs indefinitely on Windows (#328); whisper vocabularies with half-serialized special tokens now fail loudly instead of producing wrong ids (#322).

Add GigaAM-v3 speech-to-text

ai-sage/GigaAM-v3, new as the gigaam backend in v0.8.25 — the only backend added this release. Russian-only, Rotary Conformer encoder, 12 GGUFs from cstr/gigaam-v3-GGUF (CTC and RNN-T heads, each character-level or SPM "e2e", three quants each).

The e2e revisions are listed first because they emit punctuation and capitalisation natively. Verified on Russian test audio at 27× realtime:

model	output
e2e-rnnt-q8_0	"Привет." / "Это тест распознавания" / "русской речи."
ctc-q4_k	"привет это тест распознавания русской речи"

Both produce valid timestamped SRT under SE's default --max-len 50 --split-on-punct. The language list holds Russian alone — -l ru is accepted, anything else is ignored with a warning.

Add OmniVoice TTS (CrispASR)

Runs OmniVoice through the omnivoice backend, alongside the existing standalone omnivoice-tts engine — the same split SE already has between Qwen3TtsCpp and Qwen3TtsCrispAsr. It reuses the CrispASR runtime a user may already have for speech-to-text and runs as a persistent server, so the model loads once instead of once per line.

Three quants from cstr/omnivoice-GGUF plus the tokenizer companion, staged with size + SHA-256 verification (all four hashes match the HF LFS oids). Unlike VoxCPM2/MOSS-TTS this backend has a usable built-in voice, so the combo offers "Default" first and cloning is opt-in; reference WAVs are seeded from the standalone engine's voices folder.

⚠ The tokenizer is pinned to the F16 build

omnivoice-tokenizer-q8_0.gguf loads and synthesises fine, but its encoder tensors cannot be read back as f32 (read_tensor_f32: unsupported type 8 per quantizer), so encoding a reference voice yields garbage codes and the clone comes out as noise. Worse, crispasr caches the encoded reference by audio content, so a single q8_0 run poisons later F16 runs. Measured by median F0 of reference vs clone:

run	reference	clone
female ref, F16 tokenizer	191.5 Hz	190.6 Hz — intelligible, correct text
male ref, F16 tokenizer	110.2 Hz	111.8 Hz — intelligible, correct text
same, q8_0 tokenizer	110.2 Hz	481 Hz — transcribes to "I." (noise)

SE therefore only ever downloads the F16 tokenizer and pins it explicitly via --codec-model rather than leaving it to sibling discovery. A test pins that.

The exact server command line SE builds — server mode, --voice-dir, provenance flags, cloning reference — was run end to end, and its output round-trips back through sensevoice as the right words in the right voice (187.2 Hz against a 191.5 Hz reference).

Testing

dotnet test tests/UI on the merge of this branch with current main: 1127 passed, 2 failed. Both failures (SubtitleGridScrollPerformanceTests.HomeAndEnd_RealizeOnlyAViewportOfRows and SyntaxTextDocumentTests.InsertWithLineBreaksSplitsTheLine) reproduce identically on plain origin/main and are unrelated to these changes. This branch adds 17 tests, all passing.

Note: change-log.txt is deliberately untouched.

🤖 Generated with Claude Code