audiopipe

Fast speech-to-text in Rust. No Python, no FFmpeg.

Supports multiple backends — ONNX Runtime, GGML, MLX, whisper.cpp — with hardware acceleration via Metal (macOS), Vulkan, CUDA, CoreML, DirectML, or CPU.

use audiopipe::{Model, TranscribeOptions};

let mut model = Model::from_pretrained("parakeet-tdt-0.6b-v3")?;
let result = model.transcribe(&audio_f32, 16000, TranscribeOptions::default())?;
println!("{}", result.text);

Models

Model	Backend	Params	Languages	Notes
parakeet-tdt-0.6b-v3	ONNX	0.6B	25	NVIDIA Parakeet TDT. CoreML on macOS, CPU int8 fallback
parakeet-tdt-0.6b-v3-mlx	MLX	0.6B	25	Parakeet on Apple GPU. 8x faster than Whisper
qwen3-asr-0.6b-antirez	Pure C	0.6B	6000+	Fastest on CPU. BLAS + AVX2/NEON
qwen3-asr-0.6b-ggml	GGML	0.6B	6000+	Metal/Vulkan/CUDA GPU support
qwen3-asr-0.6b	ONNX	0.6B	6000+	CoreML/DirectML/CUDA via ONNX Runtime
whisper-*	whisper.cpp	varies	99	e.g. whisper-large-v3-turbo

All models auto-download from HuggingFace on first use.

Install

[dependencies]
audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["parakeet"] }

Usage

Parakeet MLX (Apple Silicon GPU — fastest)

8x faster than Whisper on Apple Silicon. Uses Metal GPU via Apple's MLX framework.

audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["parakeet-mlx"] }

Parakeet ONNX (cross-platform)

audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["parakeet"] }

Qwen3-ASR antirez (fastest CPU)

Pure C implementation by antirez. Faster-than-realtime on low-end CPUs with BLAS and custom AVX2/NEON kernels.

audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["qwen3-asr-antirez"] }

# With OpenBLAS (~10x faster matmul on Windows/Linux)
audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["qwen3-asr-antirez-blas"] }

macOS uses Accelerate automatically. For Windows OpenBLAS setup:

Invoke-WebRequest "https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.31/OpenBLAS-0.3.31-x64.zip" -OutFile OpenBLAS.zip
Expand-Archive OpenBLAS.zip -DestinationPath C:\OpenBLAS
[System.Environment]::SetEnvironmentVariable("OPENBLAS_PATH", "C:\OpenBLAS\win64", "User")

Qwen3-ASR GGML (GPU-accelerated)

Best balance of speed, accuracy, and language coverage.

# CPU
audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["qwen3-asr-ggml"] }

# macOS Metal GPU
audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["qwen3-asr-ggml", "metal"] }

# Windows/Linux Vulkan GPU
audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["qwen3-asr-ggml", "vulkan-ggml"] }

# NVIDIA CUDA GPU
audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["qwen3-asr-ggml", "cuda-ggml"] }

Qwen3-ASR ONNX

audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["qwen3-asr"] }

# With CoreML (macOS) or DirectML (Windows)
audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["qwen3-asr", "coreml"] }
audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["qwen3-asr", "directml"] }

Whisper

audiopipe = { git = "https://github.com/screenpipe/audiopipe.git", features = ["whisper"] }

Feature flags

Models

Feature	Description
parakeet-mlx	Parakeet TDT via MLX (Apple Silicon GPU)
parakeet	Parakeet TDT via ONNX Runtime
qwen3-asr-antirez	Qwen3-ASR pure C (fastest CPU)
qwen3-asr-antirez-blas	+ OpenBLAS for ~10x faster matmul
qwen3-asr-ggml	Qwen3-ASR via GGML
qwen3-asr	Qwen3-ASR via ONNX Runtime
whisper	Whisper via whisper.cpp

GPU acceleration

Feature	Description
metal	Metal GPU for GGML models (macOS)
vulkan-ggml	Vulkan GPU for GGML models (Windows/Linux)
cuda-ggml	CUDA GPU for GGML models (NVIDIA)
coreml	CoreML for ONNX models (macOS)
directml	DirectML for ONNX models (Windows)
cuda	CUDA for ONNX models (NVIDIA)

Examples

# Transcribe with Parakeet
cargo run --release --example transcribe --features parakeet -- audio.wav

# Transcribe with GGML Qwen3-ASR
cargo run --release --example transcribe --features qwen3-asr-ggml -- --model qwen3-asr-0.6b-ggml audio.wav

# Benchmark antirez (pure C, fast CPU)
cargo run --release --example bench_asr --features qwen3-asr-antirez -- audio.wav

# Benchmark with OpenBLAS
cargo run --release --example bench_asr --features qwen3-asr-antirez-blas -- audio.wav

Used by

• screenpipe — AI that knows everything you've seen, said, or heard

License

MIT