Parakeet.cpp – Parakeet ASR inference in pure C++ with Metal GPU acceleration
github.com
February 27, 2026
5 min read
Summary
Frikallo/parakeet.cpp provides an ultra-fast and portable implementation of NVIDIA's Parakeet models for on-device speech recognition in C++. It achieves approximately 27ms encoder inference on Apple Silicon GPUs for 10 seconds of audio, making it 96 times faster than CPU processing, and utilizes the Axiom tensor library for automatic Metal GPU acceleration without heavy dependencies.
Key Takeaways
- The Parakeet implementation in C++ enables fast speech recognition with NVIDIA's models, achieving approximately 27ms encoder inference on Apple Silicon GPUs for 10 seconds of audio using a 110M model.
- The framework utilizes Axiom, a lightweight tensor library, with automatic Metal GPU acceleration, eliminating the need for ONNX or Python runtimes.
- Parakeet supports various models for different tasks, including offline and streaming speech recognition, multilingual capabilities, and speaker diarization for up to four speakers.
- The system processes audio through a consistent pipeline, converting 16kHz mono WAV files into 80-bin Mel spectrograms for efficient transcription.
Community Sentiment
MixedPositives
- The ability to run multiple model families for various tasks like offline transcription and speaker diarization showcases the versatility of the Parakeet ASR engine, which is crucial for diverse applications.
- Local inference capabilities in C++ are becoming more accessible, enabling developers to create efficient pipelines for new models quickly, which can significantly enhance the development of speech recognition technologies.
- Users report being impressed with the performance of Parakeet on both Windows and Mac, indicating its reliability and effectiveness across platforms.
Concerns
- Concerns about achieving low latency under 100ms for speech recognition persist, which is critical for real-time applications and user experience.
- Some users suggest that using CoreML with the Apple Neural Engine may provide better speed and power efficiency compared to Metal, raising questions about the optimal platform for ASR inference.
Related Articles
Pure C, CPU-only inference with Mistral Voxtral Realtime 4B speech to text model
Feb 10, 2026
Rust implementation of Mistral's Voxtral Mini 4B Realtime runs in your browser
Feb 10, 2026
Flash-MoE: Running a 397B Parameter Model on a Laptop
Mar 22, 2026
Nvidia PersonaPlex 7B on Apple Silicon: Full-Duplex Speech-to-Speech in Swift
Mar 5, 2026
A CPU that runs entirely on GPU
Mar 4, 2026
Source
github.com
Published
February 27, 2026
Reading Time
5 minutes
Relevance Score
51/100
Why It Matters
This page is optimized for focused reading: quick context up top, a clean summary block, and a direct path to the original source when you want the full story.