SAVA
streaming voice activity and end-of-turn detection model
2 min readupdated
SAVA is an open-source 25.5M-parameter model from Inworld that analyzes an audio stream in real time and outputs two probabilities: whether speech is present in the frame and whether the speaker has finished their turn (end-of-turn). Its architecture combines Mamba-2 blocks with attention mechanisms, letting it cleanly separate human speech from coughs, laughter, or keyboard noise. The model runs on both CPU and GPU with minimal resource overhead, making it an ideal fit for voice assistant developers.
Key features
- Hybrid architecture built on Mamba-2 blocks and attention mechanisms.
- Low latency: voice activity detection with a 120 ms lookahead, and instant end-of-turn detection.
- Lightweight: just 25.5 million parameters, running efficiently on both CPU and GPU.
- Noise-resistant: filters out background sounds like coughs, music, and keystrokes.
What you can do
- Integrate it into a voice bot to naturally detect when a user has finished speaking.
- Run a local speech detection and segmentation pipeline using the official library.
- Build an interactive voice app with live microphone support and streaming audio processing.
Example
from inworld import sava
model = sava.load("inworld/sava")
curves = sava.run(model, sava.load_audio("recording.wav"))
turns = sava.segment(curves, sava.Policy(), model.lookahead_frames)
Read more
DetailsBenchmarks: compared against Silero VAD and pipecat smart-turn v3.2
Hardware: tested on NVIDIA GPUs and CPUs (3.8 ms per 40 ms frame on an Apple M3 Max)
Input: 16 kHz mono audio, outputting predictions every 40 ms
License: Apache-2.0, open weights and code










Comments