voxora-minimax

voxora-minimax is the first hosted-API engine adapter in the voxora workspace. It implements voxora_traits::AsrEngine against the MiniMax /v1/speech_to_text API.

Unlike voxora-whisper and voxora-qwen3asr (which load a model file from disk and run inference locally), MiniMax runs the inference server-side. The engine wraps the bearer-token-authenticated multipart/form-data upload, sends a single POST per transcription, and parses the verbose_json response (which carries the full text plus per-segment timestamps and speaker labels).

Loading the engine

use voxora_minimax::{MiniMaxConfig, MiniMaxEngine};
use voxora_traits::{AsrEngine, TranscribeOptions};

# fn run() -> Result<(), voxora_traits::AsrError> {
let config = MiniMaxConfig::new("sk-...")?;
let engine = MiniMaxEngine::new(config)?;

let samples: Vec<f32> = vec![0.0; 16_000]; // 1 s of silence @ 16 kHz
let result = engine.transcribe(&samples, &TranscribeOptions::default())?;
println!("{}", result.text);
# Ok(()) }

Authentication

MiniMax uses bearer-token auth. The token is resolved via voxora_config::VoxoraConfig::minimax_api_key:

1. Explicit value on MiniMaxConfig::new(api_key) — preferred for library callers; the key never enters process env.
2. VOXORA_MINIMAX_API_KEY env var.
3. MINIMAX_API_KEY env var (canonical alias).

The token is wrapped in a SecretString (zero-on-drop, Debug-redacted).

Sample-rate contract

Soft contract. The engine accepts any sample rate / channel layout from the caller and wraps the input &[f32] in an in-memory WAV (f32 LE PCM, mono at the engine's assumed 16 kHz). MiniMax docs recommend mono 16 kHz; high-sample-rate stereo uploads eat the 50 MB / 500 s server caps without improving transcription quality. The engine does not enforce the rate — the caller is responsible for downsampling / channel mixing.

Server caps

• 500 s audio duration (server returns 400 bad_request_error past this)
• 50 MB body size (server returns 413 invalid_request_error past this)
• 20 supported BCP-47 language tags (server returns 400 for anything outside)
• One model today: asr-1.0

Feature flags

No GPU backend features (compute is server-side). No hf feature (MiniMax has no on-disk model).

CLI

MINIMAX_API_KEY=<key> cargo run -p voxora-cli --release --features minimax -- minimax audio.wav

(Requires voxora-cli to be built with the minimax feature, which re-exports voxora-minimax via voxora-bridge.)

Live integration test

MINIMAX_API_KEY=<key> cargo test -p voxora-minimax -- --ignored live_transcription_works

The test posts 1 s of 440 Hz sine at 16 kHz mono to the real MiniMax endpoint and asserts the round-trip succeeds. MiniMax does not guarantee a specific text output for non-speech audio; the contract is just that the response is parseable.

Reference

• MiniMax API docs
• voxora-traits — the trait surface this engine implements.
• voxora-bridge — the umbrella crate that re-exports this engine behind the minimax feature flag.

License

Apache-2.0.