Looking for the AI employee system? This page documents the pre-alpha pip install ragleap-voice package, which manages speech models only. To install and run ragleap-core (AI employees, channels, Docker), see the documentation.

Pre-alpha · 88 test functions, fake engines only (no models)

ragleap-voice

Open-source (MIT), local-first speech tooling with Indian languages as a priority. Version 0.0.1 is the model-management layer only: a licence-checked manifest and hash-verified model downloads. There are no speech engines in this release, so you cannot transcribe or synthesize audio with it yet.

0.0.1

PyPI version

Runtime dependencies

Test functions

Install

Python 3.10 or newer. The core is standard library only. Nothing uses the network unless you run models pull or models lock.

pip install ragleap-voice
# or, with uv
uv add ragleap-voice

ragleap-voice models list
ragleap-voice models show indic-parler-tts
ragleap-voice models check

Commands

models list / show NAME

Read a manifest of candidate models with their licences, pinned revisions and status. Every entry is currently unverified: a research lead, not an endorsement. None is a default.

models check

Rejects verified or default entries that lack a matching lock, use licences outside the allowlist, or list pickle-capable files (formats that can run code when loaded).

models verify

Checks a cache folder against the lock files (the size and SHA-256 of every file at a pinned revision). Offline and read-only.

models pull NAME

Downloads from Hugging Face only: HTTPS, a host allowlist, and the access token (HF_TOKEN) is sent only to huggingface.co. Every file is checked against its lock before anything enters the cache.

models lock NAME

A maintainer tool that builds a lock by downloading and hashing every file.

Models

Two entries are locked so far: indic-parler-tts (a gated Hugging Face model: accept its conditions on the website and set HF_TOKEN) and its tokenizer dependency flan-t5-large-tokenizer. Both are pinned but have not had a human review, so pulling them needs --allow-unverified. Models are separate downloads with their own licences, listed in the manifest and in the third-party notices.

Measurements

Reproducible benchmarks use small samples of read speech on one GPU. In them an open Indian-language recogniser (IndicConformer) had clearly lower word error than Whisper large-v3 on Hindi, Tamil, Telugu, Malayalam and Kannada. Naturalness, emotion quality and comparisons with commercial services have not been measured, and this project makes no claim about them.

Known limitations

No speech engines. This release cannot transcribe or synthesize audio.
Every manifest entry is unverified. The two locked models are pinned but have had no human review.
Planned, not available: a speech-to-text and text-to-speech API with engines, streaming, voice profiles with a consent record, emotion and style control, a registry of community-trained models, and more languages.