Inworld

SAVA

Semantic acoustic voice activity: streaming speech and end-of-turn detection in one model with an open license.

GitHub: inworld-ai/sava


TL;DR

SAVA returns two probabilities for a live audio stream:

  • speech β†’ whether a frame holds speech
  • turn β†’ whether the speaker's turn has ended

The model's shared network uses acoustic and sentence context to distinguish speech from coughs, laughs, music and keyboard clatter.

SAVA
made for voice activity and conversational turn taking
input 16 kHz mono audio
output two probabilities every 40 ms
timing speech probability for a frame is emitted 120 ms later (lookahead); end-of-turn probability has no lookahead
parameters 25.5M
execution plain PyTorch, CPU or an explicitly selected device

Use it to open a turn when speech starts and close it when the sentence is complete. The reference turn policy turns the probabilities into segments; applications can supply their own policy. SAVA responds to any voice. It does not pick out a target speaker, and it leaves interruptions and backchannels to the application's dialogue policy.

Quick start

Using a coding agent? Point it at the setup section of AGENTS.md: it installs the package, checks the machine with sava-doctor and runs SAVA on a first recording.

With Python 3.10 or later:

python -m pip install inworld-sava
# Or from a clone: git clone https://github.com/inworld-ai/sava && cd sava && python -m pip install -e .
sava run recording.wav
sava run recording.wav --turns_dir turns/ --curves curves.json
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)

session = sava.Session(model)
for piece in pieces_of_16k_audio:  # One-dimensional floating-point mono samples.
    frames = session.push(piece)
frames = session.flush()

sava.load("inworld/sava") and the CLI's default model download the weights on first use and verify their checksum. The CLI defaults to CPU; --device selects another device. For microphone input, install python -m pip install "inworld-sava[mic]" and run sava mic.

The code, the integration guide and the benchmarks are on GitHub: inworld-ai/sava.

Hardware

Tested on NVIDIA Ampere, Hopper and Blackwell GPUs and on CPU. On an Apple M3 Max CPU one stream takes 3.8 ms per 40 ms frame; the integration guide lists the measured configurations.

The benchmarks

Voice-activity frame AUC for SAVA and Silero VAD on nine public dataset readings

End-of-turn clip AUC for SAVA and smart-turn on public datasets

Voice activity is compared with Silero VAD; end of turn is compared with pipecat smart-turn v3.2. These are per-dataset measurements, with no combined ranking. Speech scores use each model's own output timing, so SAVA's 120 ms lookahead makes the comparison unequal in latency. Quiet speech-like noise can trigger SAVA; the ESC-50 results measure that failure mode.

The complete tables, confidence intervals, noise tests, turn latency and reproduction steps are on GitHub.

Architecture

SAVA: causal log-mel features, frame mixer, Mamba and attention trunk, speech and turn heads

part detail
front end causal log-mel, 128 Slaney mel bins of a 25 ms Hann DFT every 10 ms, log10 with a 1e-10 floor, (x + 4) / 4; four hops per frame plus the frame's log RMS
mixer the current and previous frame's features, one linear projection to 512, RMSNorm
trunk three Mamba-2 blocks (expand 2, head dim 64, state 128, conv kernel 4) and one attention block (8 query heads, 2 key/value heads, head dim 64, 512-frame causal window, rotary positions, RMS-normed q and k, exclusive self attention), each with a SwiGLU FFN of 2048
voice-activity head RMSNorm, down-projection to 512, one attention block (window 512, FFN 1536), RMSNorm, two linear readouts: on time, and the released three-frame look-ahead
end-of-turn head the trunk output with a 64-wide embedding of the previous frame's on-time speech probability, the same block shape, a linear readout with four horizon columns (0, 200, 400, 800 ms); column 0 is the released output

Training data

SAVA was trained on approximately 5,000 hours of unique audio. Inworld built training examples from real and synthetic conversational speech mixed with non-verbal vocal sounds, noise, music and room acoustics, with a speech label and an end-of-turn label for every frame. The training recipe, code and data are not part of this release.

Third-party datasets used in training are credited in NOTICE.

License

Code and weights are Apache-2.0 (LICENSE). Third-party notices are in NOTICE.

How to contribute

Everyone is welcome to use the model however they like and to make it better in any way they see fit! Development happens on GitHub: see CONTRIBUTING.md for setup, tests and bug reports.


Built for people with ❀️ by Inworld.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Space using inworld/sava 1