Moonshine Streaming Tiny
Model Overview
Moonshine is a high-efficiency automatic speech recognition (ASR) model designed specifically for real-time speech recognition. Unlike Whisper, which processes audio in fixed 30-second chunks, Moonshine uses a variable-length architecture that only computes the actual duration of the speech received.
Useful Sensors developed Moonshine and released the English model as open-source. There are 2 models of different sizes and capabilities - base and tiny. The tiny version utilizes 27M parameters.
Moonshine Streaming Tiny is based on Moonshine V2, which introduced sliding window attention in the encoder. The encoder can now stream encoding frames to the decoder, so there is no need to wait for the full length of speech before emitting the first token, reducing time-to-first-token (TTFT).
Moonshine Streaming Tiny has been optimized for the Synaptics Astra® SL2610-Series processors with Torq NPU.
Model Features
- Model Type: Automatic Speech Recognition
- Input: Raw waveform (1D array of floats) 16kHz mono audio up to 30 seconds
- Output: Sequence of token IDs (integers)
- Quantization: None
Performance
Torq v2.1.0 on SL2610, mean (global average).
| Stage | Real-Time Factor (RTF) | Tokens / s |
|---|---|---|
| Moonshine Tiny Encoder | 0.38 | N/A |
| Moonshine Tiny Decoder | 0.68 | 36.6 |
Deployment
The compiled model files are available for download on Huggingface at Synaptics/moonshine-streaming-tiny-torq.
Usage Tutorial
A usage example is provided at Torq Examples / moonshine_streaming. Check out the README for instructions.
License
Both the source model and the compiled model for on-device deployment are licensed under MIT License.
Learn More
- Synaptics AI Developer Zone: Get started with documentation, tutorials and resources for your Edge AI journey.
- Astra Support Portal: Connect with our engineering team and community.
