RagLeapragleap-voice 0.0.2: still the model-management layer only.
There are no speech engines in this release. It cannot transcribe or synthesize yet.
What changed since 0.0.1:
- The Whisper large-v3 model entry now names the repository we actually run (Systran/faster-whisper-large-v3) at a pinned revision, with a lock file. The lock allows one justified exception to the pickle-format rule: a CTranslate2
model.binmarked with a format tag, header-checked, in addition to size and SHA-256. All models are still unverified and none is a default. models lock --format PATH=NAMEwrites that tag.- A first language coverage matrix as data (
coverage/stt.csv,coverage/tts.csv) with a standard-library checker. Rows come only from the measurements inbenchmarks/voice/README.md, and unmeasured languages say so. - Number-normalizer improvements for the benchmark metrics, contributed from outside the project (see the pull requests and the changelog).
- Project links point at the ragleap organisation.
Install:
pip install ragleap-voice
uvx ragleap-voice models list
Read before quoting anything: measurements are on read speech with small samples and one GPU, and every limit is stated in benchmarks/voice/README.md. Two fine-tuned African-language candidates were measured but are possibly contaminated by their training data and are not in the manifest. Nothing here is a claim about any commercial service. No security or legal review has happened yet.
Planned for 0.1.0: transcribe() and synthesize() with engines as optional extras. The design draft is in packages/ragleap-voice/docs/design/api.md.