orionq-release-rc
quantized voice activity detection model
1 min readupdated
orionq-release-rc (release candidate gptq4-g128-b03) is a lightweight Voice Activity Detection (VAD) model optimized for on-device execution. Built on the Orion-VAD 2.1 architecture, it uses 4-bit weight quantization (a model compression technique that saves memory and speeds up execution) with a group size of 128. It is designed to isolate speech segments efficiently with minimal resource overhead, making it a strong fit for embedded systems developers and local voice assistants.
Key features
- Architecture: built on Orion-VAD 2.1 with 4-bit weight compression.
- Performance: achieves a p99 latency of just 206 ms on CPUs with Intel NPU.
- Accuracy: maintains a 92.6% score on the challenging vad-eval-hard-2k benchmark dataset.
What you can do
- Embed the model into local apps for speech detection and silence filtering.
- Run it on Intel NPU-enabled devices for offline, real-time audio stream processing.
Example
The model takes an audio stream and outputs timestamps for active speech segments. Runtime configuration is stored in config.json alongside the quantized weights in weights.bin.
Read more
DetailsAccuracy (vad-eval-hard-2k): 92.6%
p99 latency (cpu-intel-npu): 206 ms
Weight precision: 4 bits, group size 128
Base model: Orion-VAD 2.1










Comments