Moonshine v2
Ultra-lightweight speech recognition engine optimized for edge devices, offering 5x faster performance compared to Whisper.
Moonshine v2 is an open-source automatic speech recognition (ASR) engine released in February 2026 by Useful Sensors, a US-based company. In a nutshell, it can be described as "Whisper's lightweight sibling – an STT that runs in real-time on Raspberry Pi, wearables, and mobile devices."
The existing OpenAI Whisper is accurate but heavy (even the tiny model is 39M+), making real-time processing difficult on edge devices. Moonshine v2 uses an "Ergodic Streaming Encoder" architecture with sliding-window self-attention, achieving 5x faster inference speed while maintaining Whisper-level accuracy with 27M parameters. It also minimizes memory and computational waste by removing zero-padding with RoPE (Rotary Position Embedding). In essence, "if Whisper is an SUV for cloud GPUs, then Moonshine is an electric bicycle for edge IoT."
From the perspective of biomedical researchers, it enables (1) real-time local transcription of patient interviews and clinical trial meeting minutes without cloud transfer (HIPAA/GDPR protection), (2) automatic storage of voice memos such as "drug treatment in 3 minutes" with timestamps during live-cell imaging, and (3) the creation of a Raspberry Pi-based experimental notebook auxiliary device. The moonshine-voice SDK provides integrated STT + intent recognition + TTS, allowing the creation of voice agents that can perform tasks like "hands-free voice commands for experiments.
💻 System Requirements
Minimum 2GB (Tiny 27M model), recommended 4GB (Base 61M model), 8GB+ (for multi-channel or Korean fine-tuning)
0 (CPU-only mode available, real-time performance from Raspberry Pi 4) / 10x acceleration with NVIDIA GPU 2GB+ / Apple Silicon Metal support
Tiny ~26MB / Base ~61MB / Total package size under 250MB. Training data is separate.
⚡ Installation
pip install moonshine-voice
Python API — Basic Transcriptionimport moonshine text = moonshine.transcribe("audio.wav", model="moonshine/tiny") print(text)
Real-time Streaming (Microphone Input)from moonshine_voice import StreamingTranscriber stt = StreamingTranscriber(model="moonshine/base") for chunk in stt.stream_from_microphone(): print(chunk.text, end="", flush=True)
Edge Device (Raspberry Pi 5) requirements: Python 3.10+, sounddevice, numpypip install moonshine-voice[edge] python -m moonshine_voice.pi --model tiny --language ko
🧬 Bio Use Cases
On-device STT for Clinical Interviews and Medical Records
Deploy the Moonshine tiny (27M) model on a Mac mini M2 to transcribe patient interviews in real-time. Zero cloud transmission (HIPAA/GDPR compliant), with a 30-minute interview taking approximately 6 seconds to transcribe. By adding intent recognition rules such as "Patient ID" and "Symptom onset time" to the moonshine-voice SDK, structured EMR input can be automatically generated.
Voice Annotation for Live-Cell Imaging
Place a Moonshine + microphone module next to a Raspberry Pi 5 microscope. Voice memos such as "Treat with 100ng/mL EGF in 3 minutes" are automatically recorded with timestamps as .jsonl logs, which are then absorbed by an ImageJ/Fiji macro as time-synchronized metadata. This can be immediately applied to live experiments where hands-free operation is required.
Edge Voice Agent (Laboratory Assistant)
Build a laboratory assistant using moonshine-voice SDK + Llama.cpp + Pi 4 (4GB). Recognize commands such as "Turn on the alarm when OD600 reaches 0.6" or "Show me the PCR master mix recipe" with latency of less than 100ms, and trigger laboratory equipment via MQTT. Zero cloud dependency, power consumption less than 5W.
FAQ
What is Moonshine v2?
Moonshine v2 is an open-source automatic speech recognition (ASR) engine released in February 2026 by Useful Sensors, a US-based company. In a nutshell, it can be described as "Whisper's lightweight sibling – an STT that runs in real-time on Raspberry Pi, wearables, and mobile devices." The existing OpenAI Whisper is accurate but heavy (even the tiny model is 39M+), making real-time processing difficult on edge devices. Moonshine v2 uses an "Ergodic Streaming Encoder" architecture with sliding-window self-attention, achieving 5x faster inference speed while maintaining Whisper-level accuracy with 27M parameters. It also minimizes memory and computational waste by removing zero-padding with RoPE (Rotary Position Embedding). In essence, "if Whisper is an SUV for cloud GPUs, then Moonshine is an electric bicycle for edge IoT." From the perspective of biomedical researchers, it enables (1) real-time local transcription of patient interviews and clinical trial meeting minutes without cloud transfer (HIPAA/GDPR protection), (2) automatic storage of voice memos such as "drug treatment in 3 minutes" with timestamps during live-cell imaging, and (3) the creation of a Raspberry Pi-based experimental notebook auxiliary device. The moonshine-voice SDK provides integrated STT + intent recognition + TTS, allowing the creation of voice agents that can perform tasks like "hands-free voice commands for experiments.
When should I use Moonshine v2?
Ultra-lightweight speech recognition engine optimized for edge devices, offering 5x faster performance compared to Whisper.
What is a biomedical use case for Moonshine v2?
On-device STT for Clinical Interviews and Medical Records: Deploy the Moonshine tiny (27M) model on a Mac mini M2 to transcribe patient interviews in real-time. Zero cloud transmission (HIPAA/GDPR compliant), with a 30-minute interview taking approximately 6 seconds to transcribe. By adding intent recognition rules such as "Patient ID" and "Symptom onset time" to the moonshine-voice SDK, structured EMR input can be automatically generated.
📝 Update Notes
No update notes yet.
🧪 Related Code of Life
No related Code of Life posts yet.