Vokra

English | 日本語

Vokra is a speech-first inference runtime written in Rust. It implements the audio pieces that general-purpose graph runtimes often leave to application code: streaming state, STFT/iSTFT and mel frontends, vocoders, neural codecs, CTC/RNN-T decoding, VAD, speaker features, pitch extraction, and audio enhancement.

Vokra loads provenance-aware GGUF files and does not load ONNX graphs at runtime. The default runtime has no third-party Cargo dependencies: the root Cargo.lock contains only first-party vokra-* crates.

Published release candidate (2026-10-18 JST): v0.4.0-rc.1 is published as GitHub release 416018534 from exact source/tag d84d311e98d0abb2987b5c4e96157ae8f2068e1a (draft=false, prerelease=true). Latest remains v0.3.0; the release has 22 uploaded assets. Final hosted run 39062234813 passed all 10 jobs: three-OS producers, fresh transferred native CLI --help and C ABI exact-version checks, Linux assembled-CLI Whisper/Silero functional smoke, and publication byte round-trip. The recovered receipt SHA-256 is d4b4656fbfa67c8f10b7dd5f6642fc896780ab11da4bc4db80d7922bb2348697, binding the Linux CLI 2a5d4ba5ca543b749f6738bd92d59da0e884a6693ab7e222b8b38ce81a91d8c4 (13,344,272 bytes) and manifest 32b03f537f48ff6f2bb7b12fda1a9426090debbd59da361c0ffa59e8f7a60e92. This remains a CPU-only preview: it does not claim macOS/Windows model inference, Apple Metal/no-fallback, full-catalog completion, performance or numerical parity, and it does not update model/Space/HF/registry/X/Scaleway artifacts.

Historical prepublication note (2026-10-17): v0.4.0-rc.1 packaging was in progress; the published Latest remains v0.3.0. At source 9ee5834d, Whisper base ASR and Silero VAD v5 speech/silence passed hosted Linux CPU functional smoke. This is not numerical parity or Apple/Metal validation. See the actual CPU results and candidate acceptance plan. The candidate installation guide is staged for publication, not a claim that its download assets already exist.

Historical campaign snapshot (2026-10-17; superseded by the published RC): packaging baseline main is fab4638a. PR #255's model-free CPU/API and normal acceptance are complete. PR #259 is closed without merge; its additional observer patch has no accepted upstream compilation/API result. Six other PRs remain open at this snapshot. The latest retained inventory is the 2026-10-16 snapshot: 194 repositories, 136 code/artifact-full and 58 non-full; it is not a fresh inventory or an Apple hardware pass count. Per-model non-Apple prerequisites, owner/legal decisions, independent real-weight CPU, final packets and Apple CPU/Metal/no-fallback remain incomplete. See the current status and remaining tasks.

Historical audit snapshot (2026-10-13): the workspace is 0.3.0 development; the latest VibeVoice Realtime-0.5B closure audit was a static-only VAST audit on disposable worker 55648877. It authenticated the selected source, lock, collector, and provenance inputs, then verified recovery and destruction of the worker. It did not download or execute model weights, run a model, invoke a package/native API, or upload an artifact. The collector verdict is OWNER_REVIEW_REQUIRED_NO_UPLOAD; the controller completed separately. Owner/legal approval, license review, provenance review, native runtime status, and runtime parity are unresolved. The audit found missing bundled license files for safetensors 0.5.3 and tokenizers 0.22.2, plus ten native packages requiring review. See the VibeVoice static-closure audit and documentation refresh record. At exact main HEAD 7d38061c3dcb56f6a16bcf4b8085d966b3cdab09, the fresh metadata/card-only audit reports 194 repositories, 193 GGUF-bearing repositories, 198 GGUF files, CPU 136/43/14/1, and Metal 136/57/1, leaving 58 unresolved public rows. This audit does not claim Apple CPU/Metal/no-fallback verification.

Historical catalog snapshot (2026-10-08): the workspace is 0.3.0 development; this documentation review and the read-only public audit use exact GitHub main checkout d100d93778191ccab77bd1c37fe3552e3d889758. The metadata/card-only public audit on 2026-10-08 reports 194 repositories, 193 GGUF-bearing repositories and 198 GGUF files: CPU code status is 136 full, 43 partial, 14 without a runtime binder and 1 non-artifact. That dated audit recorded these counts; it did not download or execute model weights. Apple Metal is 136 full, 57 blocked by CPU and 1 non-artifact. VibeVoice Realtime-0.5B now has a strict structural vibevoice_streaming binder and CLI inspection route, so it is partial; its complete 605-tensor checkpoint has been converted and a narrow independent Microsoft-source reference probe is accepted, but full streaming, native CPU parity, and Apple CPU/Metal verification remain pending. Qwen3-TTS PR #109 subsequently completed a clean, no-upload VAST run for all four public variants plus the shared 12 Hz decoder: strict official-weight reload and independent CPU parity passed 4/4, and the recovered batch closed 38/38 checksums. Apple CPU/Metal and no-fallback verification, and any public-artifact replacement, remain pending. The bounded Apple Silicon batch recorded in the Apple results passed only for its named scopes, and four explicitly approved artifacts were subsequently published. That snapshot recorded 58 unresolved public rows; refresh live metadata before treating these counts as current. On 2026-10-04, a corrected VAST source-only audit authenticated its fixed Kyutai source inputs and independently verified worker cleanup. It did not acquire or execute model weights, and it does not advance CPU parity, Apple hardware, or publication status. Version 0.3.0 is the first GitHub source release; external package registries remain disabled. Vokra remains pre-1.0, so Rust APIs, the C ABI, GGUF metadata and model coverage may change. Pin the exact v0.3.0 tag when evaluating this release in another project.

Why Vokra

• Audio-native execution: speech frontends, streaming caches, decoders, vocoders, codecs, VAD, and enhancement are native operators rather than ONNX graph glue.
• Small dependency surface: runtime crates depend only on first-party vokra-* crates. Offline conversion remains separate from runtime loading.
• Explicit failures: unsupported operations and unavailable devices return errors; GPU work never silently falls back to CPU.
• Reproducible model files: Vokra GGUF metadata records frontend settings, topology, quantization policy, source provenance, and licence information.
• Portable integration: CPU is the default; Metal, CUDA, Vulkan, and WebGPU are opt-in. A generated C header supports native and language bindings.

Quick start

You need Git and Rust 1.89 or newer (the effective MSRV; the CPU backend requires 1.89). Build the CLI from source:

git clone https://github.com/ayutaz/vokra.git
cd vokra
cargo build --release -p vokra-cli

Download the published Whisper base GGUF and run the included public-domain audio fixture:

curl -L https://huggingface.co/vokra/whisper-base/resolve/main/whisper-base.gguf \
  -o whisper-base.gguf
target/release/vokra-cli run \
  --model whisper-base.gguf \
  --input tests/fixtures/audio/jfk-30s.wav

Use the built-in help before converting or running another architecture:

target/release/vokra-cli --help
target/release/vokra-cli convert --help
target/release/vokra-cli run --help

The getting-started guide covers conversion, VAD, TTS, benchmarking, and the C ABI.

Model and backend status

Vokra covers ASR, TTS, speech-to-speech, VAD and turn-taking, keyword spotting, speaker processing, pitch, codecs and vocoders, enhancement, separation, and audio understanding. Maturity is tracked per architecture: a converter, a GGUF loader, a native forward pass, numerical parity, and a published artifact are separate milestones. The existence of one does not imply the others.

Use these sources instead of a copied model list:

• vokra-cli convert --help — accepted converter identifiers;
• vokra-cli run --help — CLI-routed inputs, outputs, and backend options;
• the Vokra model hub — published artifacts and model-specific licence cards;
• crates/vokra-cli/src/engine.rs — explicit runtime routing and deferred-operation registry for developers.

CPU is the default backend. Metal, CUDA, Vulkan, and WebGPU are opt-in and have operation-specific coverage. CoreML has an experimental whole-submodel delegate path for the Whisper encoder; QNN remains an SDK-gated experimental delegate scaffold. See the backend guide before selecting an accelerator.

The current coverage snapshot separates source/runtime readiness from numerical evidence and publication. Incomplete routes fail closed; unsupported operations never fall back silently to CPU. The UTMOS legacy Lightning checkpoint is also intentionally refused by the restricted weights_only=True loader, so no UTMOS numeric parity result is claimed.

Library integration

Build the C library with:

cargo build --release -p vokra-capi

include/vokra.h is the generated C reference. The API index links to the Rust and binding surfaces, including Python, Swift/iOS, Unity, Godot, Android, web, and server examples. The C ABI remains pre-1.0 and is not frozen.

Documentation

• Documentation map
• Getting started
• CLI tutorial
• Architecture
• Backend guide
• Migration guide
• Licence audit and legal/compliance notes

Contributing

Contributions are welcome. Read CONTRIBUTING.md before a large change, and use the good first tasks for scoped entry points. Bugs and proposals can be filed in GitHub Issues. Participation is governed by the Code of Conduct. Report vulnerabilities privately as described in the Security Policy, not in a public issue.

Licence

Vokra source code is licensed under Apache-2.0. Model weights and reference assets may use different licences; review each model card, docs/license-audit.md, and NOTICE before redistribution or commercial use. Non-commercial weights are excluded from the default publication path unless an explicit research-only gate is used.