Speech recognition overview

​

Available models

​

Performance comparison

​

Key features

​

Pure Rust implementation

• Native Metal GPU acceleration via MLX
• Efficient memory management
• Cross-platform binary distribution
• Direct integration into Rust applications

​

Optimized for Apple Silicon

• Metal GPU acceleration for neural network operations
• Accelerate framework for audio processing (FFT, resampling)
• 8-bit quantization support for reduced memory usage
• Efficient batch processing for long-form audio

​

Production-ready API

# Start API server
cargo run --release -p ominix-api -- \
    --asr-model ~/.OminiX/models/qwen3-asr-1.7b --port 8080

# Transcribe audio (OpenAI Whisper-compatible)
curl http://localhost:8080/v1/audio/transcriptions \
    -F [email protected] -F language=English

​

Architecture overview

​

Qwen3-ASR architecture

Audio (16kHz) → 128-mel Spectrogram → Conv2d×3 (8× downsample)
             → Transformer Encoder → Linear Projector → Qwen3 Decoder → Text

​

Paraformer architecture

Audio (16kHz) → 80-mel Spectrogram → LFR 7/6
             → SAN-M Encoder (50 layers) → CIF Predictor
             → Bidirectional Decoder → Tokens (parallel)

​

FunASR-Nano architecture

Audio (16kHz) → 80-mel Spectrogram → Whisper Encoder (frozen)
             → Audio Adaptor → Qwen LLM → Text

​

Supported audio formats

• WAV - Native support (any sample rate, mono/stereo)
• MP3, M4A, FLAC, OGG, AAC - Automatic conversion via ffmpeg
• Raw samples - Direct f32 array input at 16kHz

​

Model selection guide

​

Choose Qwen3-ASR when you need:

• Multilingual support (30+ languages)
• Best accuracy on Chinese, English, Japanese, Korean
• Long-form audio transcription (automatic 30s chunking)
• Production-grade quality and speed balance

​

Choose Paraformer when you need:

• Chinese-only transcription
• Maximum speed (non-autoregressive)
• Lower memory footprint
• Extremely fast inference for short audio

​

Choose FunASR-Nano when you need:

• 31 language support including dialects
• Far-field/noisy environment robustness
• Regional accent recognition
• LLM-based semantic understanding

​

Quick start

# Qwen3-ASR-1.7B (recommended)
huggingface-cli download mlx-community/Qwen3-ASR-1.7B-8bit \
    --local-dir ~/.OminiX/models/qwen3-asr-1.7b

# Paraformer
git clone https://modelscope.cn/models/damo/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch.git

# FunASR-Nano
huggingface-cli download mlx-community/Fun-ASR-Nano-2512-fp16 \
    --local-dir ~/.OminiX/models/funasr-nano

# Qwen3-ASR
cargo run --release --example transcribe -- audio.wav

# Paraformer (after conversion)
cargo run --release --example transcribe -- audio.wav /path/to/paraformer

# FunASR-Nano
cargo run --release --example transcribe -- audio.wav

use qwen3_asr_mlx::{Qwen3ASR, default_model_path};

let mut model = Qwen3ASR::load(default_model_path())?;
let text = model.transcribe_with_language("audio.wav", "English")?;
println!("Transcription: {}", text);

​

Next steps