voxora

Model-agnostic Speech-to-Text for Rust.

A candle-native bridge that unifies Whisper, Qwen3-ASR, and future Hugging Face audio models behind one trait, so any Rust application can swap engines without touching inference code.

Status: pre-alpha / engine adapters shipped. The CLI is in voxora-cli/. The investigation recap and the phased roadmap are in docs/.

Why voxora?

The name is a Latin portmanteau, in the same construction style as Telora:

vox + ora = voxora — "voice by mouth" — a name that describes what a speech-to-text engine does (turn voice into uttered text) and that sounds right when spoken, because it ends in the open vowel ah, the sound produced by opening the mouth to speak.

It was chosen from a curated list of ~60 candidates across Latin, Greek, and modern coinages; the criteria were:

• Free on both crates.io and GitHub (airvzxf/voxora).
• Phonetically pronounceable in Spanish and English.
• Mirrors the construction of Telora (tele + ora), so the brand family reads.
• No hyphens (ergonomics in use voxora::...), no -rs suffix (convention for native Rust crates), no domain suffix like -asr (the name describes the brand, not the function).

Why this project exists

Telora is a local Speech-to-Text assistant for Linux, written in Rust and originally bound to whisper-rs. As the maintainer explored alternative models (specifically Qwen3-ASR, via huggingface/candle#3509 and the qwen3-asr-rs fork), two problems surfaced:

1. candle's maintainers prioritize NVIDIA-native optimization over new model architectures. The Qwen3-ASR PR has been waiting for review for months despite CPU/GPU benchmarks that match or beat PyTorch on RTX 3090.
2. The Rust ASR ecosystem is split between ONNX Runtime and whisper.cpp. transcribe-rs (233⭐) already provides a SpeechModel trait that unifies nine engines, but every engine except whisper.cpp goes through ORT — never candle.

voxora is the candle-native sibling: one trait that wraps candle-native inference engines (Qwen3-ASR today, Voxtral/Granite-Speech tomorrow), with auto-resolution from Hugging Face and a hardware dispatcher (CUDA → Metal → CPU) that lets downstream apps pick the best available accelerator at runtime.

Telora will eventually depend on voxora instead of whisper-rs directly, becoming model-agnostic the way the maintainer originally envisioned.

How it fits in the stack

┌─────────────────────────────────────────────────────────────┐
│ Applications: Telora today, any future Rust STT consumer     │
└────────────────────────┬────────────────────────────────────┘
                         │ depends on
                         ▼
┌─────────────────────────────────────────────────────────────┐
│ **voxora** — the candle-native ASR bridge                   │
│   • AsrEngine trait, TranscribeOptions, TranscriptionResult │
│   • Hugging Face model resolution + quantization selection │
│   • Per-engine adapters (whisper-rs, qwen3-asr-rs, …)       │
│   • Hardware dispatcher: CUDA → Metal → CPU                 │
└────────────────────────┬────────────────────────────────────┘
                         │ depends on
        ┌────────────────┼────────────────┐
        ▼                ▼                ▼
   whisper-rs     qwen3-asr-rs     future engines
   (whisper.cpp   (candle, via     (Voxtral, Granite-
    bindings)      airvzxf fork)    Speech, Parakeet…)

Status and roadmap

The phased plan is in docs/ROADMAP.md. The short version:

Coordinated releases

voxora publishes a unified version across every workspace crate that participates in a release. When voxora X.Y.0 ships, all participating crates ship at X.Y.0 — you can depend on them with matching versions and trust they stay in lockstep:

voxora-traits   = "X.Y.0"
voxora-engine  = "X.Y.0"
voxora-hf      = "X.Y.0"
voxora-whisper = "X.Y.0"

Picking a single version X.Y.0 and using it across the voxora-* crates in your Cargo.toml is the supported way to consume voxora. If you depend on, say, voxora-traits = "X.Y.0" and voxora-engine = "X.Y.0", you can be confident they were released together and that their public surfaces are wire-compatible. No need to read per-crate changelogs to figure out which combinations of minor versions are compatible — the workspace version is the compatibility promise.

Upgrade guide — 0.3.x → 0.4.0. The voxora-core compatibility shim that existed in 0.3.0 / 0.3.1 is removed in 0.4.0. The trait surface has lived in voxora-traits since 0.3.0. To upgrade:

// 0.3.x
use voxora_core::AsrEngine;
use voxora_core::TranscribeOptions;

// 0.4.0
use voxora_traits::AsrEngine;
use voxora_traits::TranscribeOptions;

In your Cargo.toml, swap voxora-core = "0.3" for voxora-traits = "0.4". If you used the voxora-core/serde Cargo feature, switch to voxora-traits/serde. The voxora-bridge umbrella crate is unaffected and continues to re-export voxora-traits for one-stop consumption.

The full invariant (including the additive-change exception) is documented in AGENTS.md → Version coordination.

Quickstart

# Build:
cargo build --release -p voxora-cli

# Download a model:
./target/release/voxora download Qwen/Qwen3-ASR-0.6B

# Transcribe a WAV (engine auto-detected from config.json):
./target/release/voxora run Qwen/Qwen3-ASR-0.6B samples/jfk.wav

# Or pin a specific engine:
./target/release/voxora run ggerganov/whisper.cpp samples/jfk.wav \
    --engine whisper --language en

See voxora --help for the full surface. Engine selection falls back to a --engine <whisper|qwen3-asr> override when config.json doesn't disambiguate. Hardware flags mirror the engines (--features cpu (default), metal, cuda).

Investigation

Why this repo exists, the gap it fills, and the options we considered are documented in docs/INVESTIGATION.md. Read it before opening an issue — most "why not just X?" questions are answered there.

License

Apache License, Version 2.0. See LICENSE.

Contributing

See CONTRIBUTING.md and our Code of Conduct.