starfleet/computer
starfleet/computer
Watch
1
0
Fork
You've already forked computer
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

feat: add OmniVoice TTS backend with quantization and FlashInfer options #15

Merged
troed merged 4 commits from feat/omnivoice-tts into devel 2026-09-07 22:20:28 +02:00
Conversation 0 Commits 4 Files changed 12 +890 -50
troed commented 2026-09-07 22:20:14 +02:00
Owner
Copy link

Summary

Adds OmniVoice (k2-fsa) as a TTS backend: 600+ languages, zero-shot voice cloning and voice design, 0.6B fp16 weights + 805MB audio tokenizer — far under the Qwen Q4_K_M VRAM budget.

  • Ported handler + argument class (adapted from upstream speech-to-speech PR #527, with fork's simple EndOfResponse semantics), wired into s2s_pipeline mirroring pocket_tts (backend arg --tts omnivoice, computer[omnivoice] pip extra).
  • bitsandbytes quantization on load: --omnivoice_quant none|int8|nf4|fp4 (CUDA only); NF4 measured fastest unaccelerated (RTF ~0.19, static ~1.8 GiB).
  • Optional FlashInfer acceleration (--omnivoice_flashinfer, --omnivoice_cuda_graph): RTF ~0.05 (~4x). Requires flashinfer-python + git-source omnivoice (see TTS README). Quant+flashinfer combo is guarded (FlashInfer weight packing requires unquantized Linear).
  • Docs: TTS README section with usage, measured VRAM/RTF table, and license notes (Apache-2.0 code; weights CC-BY-NC 4.0 + Boson Higgs Community License — non-commercial).
  • Tests: 18 OmniVoice handler tests (TDD), full suite 930 passed / 16 skipped, ruff clean.
## Summary Adds OmniVoice (k2-fsa) as a TTS backend: 600+ languages, zero-shot voice cloning and voice design, 0.6B fp16 weights + 805MB audio tokenizer — far under the Qwen Q4_K_M VRAM budget. - Ported handler + argument class (adapted from upstream speech-to-speech PR #527, with fork's simple EndOfResponse semantics), wired into s2s_pipeline mirroring pocket_tts (backend arg `--tts omnivoice`, `computer[omnivoice]` pip extra). - bitsandbytes quantization on load: `--omnivoice_quant none|int8|nf4|fp4` (CUDA only); NF4 measured fastest unaccelerated (RTF ~0.19, static ~1.8 GiB). - Optional FlashInfer acceleration (`--omnivoice_flashinfer`, `--omnivoice_cuda_graph`): RTF ~0.05 (~4x). Requires flashinfer-python + git-source omnivoice (see TTS README). Quant+flashinfer combo is guarded (FlashInfer weight packing requires unquantized Linear). - Docs: TTS README section with usage, measured VRAM/RTF table, and license notes (Apache-2.0 code; weights CC-BY-NC 4.0 + Boson Higgs Community License — non-commercial). - Tests: 18 OmniVoice handler tests (TDD), full suite 930 passed / 16 skipped, ruff clean. 
Port k2-fsa OmniVoice (600+ languages, voice cloning, voice design)
as --tts omnivoice, thread its kwargs through s2s_pipeline mirroring
pocket TTS, add the omnivoice pip extra and a section in the TTS
README. Weights are CC-BY-NC 4.0 (non-commercial).
feat: guard quant+flashinfer combo, document measured VRAM/RTF
Some checks failed
CI / Sanity check (ubuntu-latest) (pull_request) Failing after 2m22s
Details
fdb092939e
troed merged commit 1a376270a2 into devel 2026-09-07 22:20:28 +02:00
troed deleted branch feat/omnivoice-tts 2026-09-07 22:20:28 +02:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
starfleet/computer!15
No description provided.