airvzxf / airvzxf/voxora

Implement voxora-voxtral engine (Mistral Voxtral ASR via Rust binding)

Open
#46 1 comment 0 reactions 0 assignees View on GitHub
blocked engine enhancement pending-engine phase-7
Dominant language
Rust
Stars
0
Forks
1
Avg merge
11m
Merged PRs (30d)
47

Description

Tracks the planned `voxora-voxtral` engine crate. Per ROADMAP
Phase 7.

## Status

**Blocked on upstream.** Mistral Voxtral (released 2025) has no
Rust binding. The reference implementation is Python (transformers
+ safetensors). A Rust port would either be from scratch, a port
of `transformers` pieces, or a candle-based reimplementation.

## What `voxora-voxtral` will be

A new workspace crate that:

- Implements `voxora_traits::AsrEngine` and
`voxora_engine::EngineAdapter` for Mistral Voxtral.
- Supports `voxora-bridge::EngineFamily::Voxtral`.
- Uses `voxora-hf` for model download.
- Supports `voxora-backend` device selection.

## Upstream blockers

- [Voxtral](https://huggingface.co/mistralai/Voxtral) is published
as safetensors weights. The inference code is in Python
(`transformers` library).
- A candle-based implementation would be the cleanest Rust path
but requires significant work.

## Acceptance

- `voxora-voxtral 0.X.0` published.
- `voxora-bridge::EngineFamily::Voxtral` added.
- Parity test against a real Voxtral checkpoint produces the
expected transcript.

## Out of scope

- Training Voxtral.
- Vision-mode of Voxtral (it can also take images). For now we
target the audio-only ASR use case.

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.