Inworld releases SAVA model for faster voice-agent turn detection

Inworld AI has released SAVA, a semantic acoustic model that combines speech activity detection with turn detection for voice applications. The company says the model can distinguish an unfinished pause from the end of a speaker's turn, helping agents respond at a more appropriate time.
- SAVA combines speech activity detection and turn detection in one model.
- It returns two probabilities every 40 ms.
- SAVA detected 85.5% of AMI meeting turn ends within 200 ms.
- The model has 25.5M parameters and runs in PyTorch.
- Developers can load its weights from Hugging Face.
One model for speech and turn detection
SAVA ships alongside Inworld's Voice Isolator and already runs turn-taking in the company's voice products. It is designed to address a common problem in conversational systems: responding too early when a person pauses to think, or waiting too long and making the interaction feel slow.
The model processes audio continuously and returns two probabilities every 40 ms. One estimates whether a frame contains speech, while the other estimates whether the speaker's turn has ended. SAVA uses the sentence so far to distinguish a pause in an unfinished thought from a completed sentence. The same context helps it separate speech from coughing, laughter, music and typing.
Testing on conversations and noisy audio
Inworld compared SAVA with Silero VAD 6.2.1 for speech detection and pipecat's smart-turn v3.2 for turn detection. The company used public datasets and the same scoring procedure for both models, with scripts, dataset manifests and raw results available in the repository.
On 5,043 turn ends in the AMI meeting corpus, SAVA detected 85.5% within 200 ms, compared with 30.8% for smart-turn. Within 400 ms, the figures were 91.2% and 41.8%. SAVA's median detection delay was 71 ms, while smart-turn's was 210 ms.
In speech detection, Inworld says SAVA had a higher frame AUC on seven of nine dataset readings against Silero. The comparison was not latency-matched because SAVA uses 120 ms of lookahead and Silero uses none. SAVA trailed on AliMeeting, which contains Mandarin meetings, and was level with Silero on AISHELL-4.
Performance and model details
Inworld reports mixed results on other turn-detection tests. On the pipecat smart-turn test set, SAVA recorded an AUC of 0.980, compared with 0.983 for smart-turn. On LiveKit's eot-bench, smart-turn led when either model could trigger during a pause, while SAVA led when predictions were read 0.2 seconds into the pause. SAVA's score on Easy Turn was 0.51, compared with 0.83 for smart-turn.
SAVA accepts 16 kHz mono audio and runs in PyTorch. It has 25.5M parameters, and Inworld says it processes one stream in under 3 ms per 40 ms frame on a single Apple M3 Max CPU thread. Its weights download from Hugging Face when the model is first loaded.
What developers need to handle
SAVA is built to supply probabilities, while a developer's turn policy decides when the agent acts on them. Inworld provides a reference policy for converting the model's output into turn boundaries, along with an integration guide covering streaming and the model's limits.
The model detects any voice it hears rather than selecting a particular speaker. Dialogue systems still need to manage interruptions and brief acknowledgments such as "mm-hm". Inworld also says SAVA is deliberately sensitive to quiet and distant speech, which means a television in another room or quiet noise resembling speech can trigger it. The company recommends Voice Isolator to remove that kind of noise first.
Inworld AI: Introducing SAVA


