v1.0.1
pathwaycom/llm-appv1.0.1May 22, 2026by KoljaB
AI Summary
Introduces major new transcription engines for Kroko/Banafo and Meta Omnilingual ASR, along with installation helpers and streaming interface improvements.
Key Highlights
- Added Kroko/Banafo ASR support with `kroko_onnx` engine and realtime previews
- Added Meta Omnilingual ASR support via the `omnilingual_asr` engine
- Introduced Kroko-ONNX installer helper (`stt-install-kroko`)
- Implemented a generic streaming transcription session interface
New Features
- Kroko/Banafo ASR engine (`kroko_onnx`)
- Meta Omnilingual ASR engine (`omnilingual_asr`)
- Kroko-ONNX installer helper
- Generic streaming transcription session interface
- Kroko realtime preview support
Full Release Notes
# RealtimeSTT v1.0.1
### Major Features
- **Kroko/Banafo ASR Support**
- Added the `kroko_onnx` transcription engine for Kroko/Banafo `.data` streaming models.
- Added engine aliases for Kroko/Banafo usage through the transcription engine system.
- Added realtime preview support for streaming engines using persistent transcription sessions.
- Kroko realtime previews feed only newly recorded audio frames instead of repeatedly reprocessing the full buffer.
- Kroko final transcription remains one-shot; the streaming path is used for realtime previews when the selected realtime engine supports it.
- **Kroko-ONNX Installer Helper**
- Added `stt-install-kroko`, exposed through the `kroko-builder` extra.
- The helper builds and installs Kroko-ONNX for the active Python environment.
- On Windows, the current Kroko builder path targets CPython 3.12 x64 and uses Docker Desktop.
- Public Kroko Community models with known filenames can be downloaded into the RealtimeSTT cache.
- **Meta Omnilingual ASR Support**
- Added the `omnilingual_asr` transcription engine for Meta Omnilingual ASR.
- Supports the published CTC and LLM model-card plumbing from the upstream `omnilingual-asr` package.
- Uses `omniASR_CTC_1B_v2` as the default Omnilingual model when the recorder is still configured with a Whisper-style default model name.
- In-memory audio is passed as predecoded waveform dictionaries to avoid upstream raw-array handling issues.
### Improvements
- **Streaming Engine Interface**
- Added a generic streaming transcription session interface.
- Engines can now opt into incremental realtime decoding.
- Existing engines keep the full-buffer fallback behavior.
- **Kroko Realtime Behavior**
- Kroko model cadence is used to choose automatic finalization tail padding.
- Added support for suppressing Kroko native stdout/stderr during recognizer calls with `suppress_native_output=True`.
- Added `KROKO_ONNX_SUPPRESS_LICENSE_OUTPUT=1` handling for compatible Kroko builds.
- **Install And Smoke-Test Guidance**
- Documented the recommended Kroko recorder install path:
`pip install "RealtimeSTT[kroko-builder,silero-onnx-cpu]"`
- Clarified that `kroko-builder` builds Kroko-ONNX, while `silero-onnx-cpu` provides the local VAD backend needed by recorder-based smoke tests and live microphone use.
- Added a public Kroko smoke script: `tests/realtimestt_kroko_test.py`.
- Added a public Omnilingual smoke script: `tests/realtimestt_omnilingual_test.py`.
- **Docs**
- Added Kroko-ONNX engine documentation.
- Added Omnilingual ASR engine documentation.
- Added installation guidance for platform-sensitive extras.
- Added `docs/licenses.md` with engine and model-family license notes.
- Clarified that `example_fastapi_server` is a source-checkout reference server, not installed by the PyPI wheel.
### Fixes
- **Omnilingual Runtime Handling**
- Fixed Omnilingual audio handoff so in-memory audio is passed in the format expected by the upstream backend.
- Added clearer dependency/platform guidance for the current Omnilingual stack.
- **Kroko Smoke Path**
- Updated Kroko smoke-test guidance so clean installs include the needed local VAD backend.
- The standalone Kroko smoke script now gives a direct hint when the recorder VAD dependency is missing.
### Compatibility Notes
- Existing default `faster_whisper` usage remains the compatibility path.
- Kroko recorder/live usage should install:
`pip install "RealtimeSTT[kroko-builder,silero-onnx-cpu]"`
- The Kroko Windows builder currently requires CPython 3.12 x64 plus Docker Desktop with the Linux engine running.
- Meta Omnilingual ASR is intended for Linux/WSL2 with Python 3.11.x.
- Native Windows Omnilingual runtime is not supported because `fairseq2n` has no Windows wheel.
- Python 3.12.x is not currently a practical Omnilingual target because of upstream `omnilingual-asr` package metadata.
- Licensed Kroko Pro models require a Pro-capable Kroko wheel and a key supplied through configuration, CLI, or environment variables.
### Other
- Added focused Kroko and Omnilingual unit coverage.
- Added public manual smoke paths for Kroko and Omnilingual.
- Package version bumped to `1.0.1`.