Introducing SAVA
Speech detection and turn detection in a single streaming model. It can run on a single CPU thread, with code and weights released under Apache 2.0.View on GitHub ↗Hugging Face ↗
Today we're also releasing SAVA, a semantic acoustic voice activity model that handles speech detection and turn detection together. It ships alongside Voice Isolator, and it already runs turn-taking in our voice products.
A voice agent that waits for silence responds at the wrong time. People pause to find a word, take a breath or work out what they want to say. Treat every pause as an ending and the agent interrupts. Wait out every pause and it feels slow. The agent also has to tell a voice from a door slamming.
01Purpose
A speech pipeline needs voice activity detection and turn detection working together, and running a separate model for each complicates the infrastructure. The open-source options for both struggle with noisy audio and support few languages well. A pause after "and after that, you…" sounds a lot like the pause after a finished sentence, so the agent either jumps in early or waits for a timer.
SAVA does both jobs in one fast model, across languages and on messy audio. It processes the audio continuously. Every 40 ms, it returns two probabilities: whether the frame contains speech and whether the speaker's turn has ended. It uses the sentence so far to tell a pause in an unfinished thought from an ending, so it can call a turn over within tens of milliseconds of the last word. The same context helps it tell speech from coughing, laughter, music and typing.
02Demo
The figure below is a short exchange between a voice agent and a caller, on two tracks. SAVA hears the caller's track, and the figure shows its output on it: the speech and end-of-turn probabilities.
Play it with sound, or drag across the recording to inspect any frame. At 5.4 s the caller stops mid-sentence for most of a second, and SAVA does not take the silence for the end of the turn. The model supplies the probabilities, and your turn policy decides when to act on them.
03Evals
We tested SAVA against Silero VAD 6.2.1 for speech detection and pipecat's smart-turn v3.2 for turn detection. Each comparison uses public datasets and the same scoring procedure for both models. The scripts, dataset manifests and raw results are in the repository.
- Inworld SAVA
- smart-turn v3.2
The largest difference was in turn detection on real conversations. We measured 5,043 turn ends in the AMI meeting corpus, using the point where one speaker stops and another takes the floor. SAVA detected 85.5% within 200 ms, and smart-turn 30.8%. The median detection delay was 71 ms for SAVA and 210 ms for smart-turn.
- Inworld SAVA
- Silero VAD 6.2.1
- Chance
For speech detection, SAVA has a higher frame AUC on seven of the nine dataset readings. On five, the difference is clear of the confidence interval. It trails on AliMeeting, which contains Mandarin meetings, and is level on AISHELL-4. On ESC-50, a dataset of everyday sounds with no speech, SAVA stays at or below 0.5 throughout 93.2% of clips, against 87.1% for Silero.
The other turn-detection results are mixed. On the pipecat smart-turn test set, the models are level: AUC 0.980 for SAVA and 0.983 for smart-turn.
On LiveKit's eot-bench, the result depends on when the prediction is read. If either model can trigger at any point during a pause, smart-turn leads, 0.86 to 0.82. At 0.2 s into the pause, the usual setup for smart-turn, SAVA leads, 0.81 to 0.78.
Easy Turn is SAVA's weakest result: 0.51, where smart-turn scores 0.83. We suspect the way its clips are built matters. Most use synthesized speech that is joined and labeled afterward, and their endings may differ from the real conversations SAVA learned from. We haven't confirmed that explanation.
See the benchmark README for the full tables, per-language results and confidence intervals.
04How to
SAVA accepts 16 kHz mono audio and runs in PyTorch. The model has 25.5M parameters. On a single Apple M3 Max CPU thread, it processes one stream in under 3 ms per 40 ms frame. The weights download from Hugging Face the first time you load it.
pip install inworld-sava
# two probabilities for every 40 ms frame of the recording
sava run recording.wav --curves curves.jsonfrom inworld import sava
model = sava.load("inworld/sava") # downloads the weights on first use
curves = sava.run(model, sava.load_audio("recording.wav"))
curves.speech # P(speech) per 40 ms frame
curves.turn # P(the turn has ended) per frame
# or stream it, in pieces of any size
session = sava.Session(model)
for piece in pieces_of_16k_audio:
frames = session.push(piece)The repository includes a reference policy for turning the model's probabilities into turn boundaries. The integration guide covers streaming and the model's limits.
SAVA detects any voice it hears. It doesn't select a particular speaker, and your dialogue policy still needs to handle interruptions and brief acknowledgments such as "mm-hm".
The model is deliberately sensitive to quiet and distant speech. That also means a television in the next room, or very quiet noise that resembles speech, can trigger it. Voice Isolator removes that kind of noise first.
05Next
If you try SAVA in a product, we'd like to hear how it performs on your audio. You can open an issue, send a pull request or talk to us on Discord. A recording of a missed turn or a false trigger is the most useful thing you can send.