Tools / Analysis
Inworld gives a voice agent two ways to listen
Voice Isolator cleans a recording. SAVA listens for the end of a thought. Inworld released code and weights for both, plus a separate report on the voice that talks back.

“The first train leaves at …” A person pauses to check the time. The pause is audible even on a perfect microphone. What a voice assistant does next depends on whether it hears the silence as the end of the speaker’s turn. Cleaning up the train announcement in the background would solve a different problem.
On November 18 Inworld released the code and weights for two models under Apache 2.0. Voice Isolator estimates which sound in a recording belongs to the voice. SAVA estimates when someone is speaking and when the turn is complete. The open release lets a team try each decision on its own audio and trace which one went wrong when an exchange feels awkward.
Listen to what the filter takes away
For recorded dialogue, Voice Isolator’s offline version can examine an eight-second chunk before deciding what to keep at any point within it. A live call cannot afford that wait. The streaming version looks just 20 milliseconds ahead. Both work at 48 kHz. Inworld puts the streaming model’s algorithmic delay at 63 milliseconds, before hardware processing. Twenty milliseconds is the amount of future audio it can consult, not a promise about when the listener will hear the result.
The demo gives a dialogue editor two things to audition: the isolated voice and the background left over after subtracting it. The second track can contain some of the voice. If a breath or the end of a word has migrated there, it is worth going back to the original before accepting a cleaner take. One input supplied in the repository, sample_14.wav, is a 10.77-second mono PCM recording at 48 kHz. That is an accessible starting point for an A/B comparison on the machines that will actually do the work.
The repository draws two practical limits. The model does not offer a control for choosing one speaker from several, and this version does not restore lost audio. An editor with two actors talking over each other still needs the separate microphone recordings if the director wants one performance alone.
Sources: Inworld / GitHub [3] · Inworld [1] · Inworld / GitHub [4]
Inworld’s original release video
Loading post from X…
The 33-second video accompanied Inworld’s announcement of Voice Isolator, SAVA and the TTS-2 report.
The silence before the time of the train
SAVA accepts 16 kHz mono audio and produces two probabilities every 40 milliseconds. One measures speech activity. The other measures whether the speaker has finished. The timing is deliberately different: the speech score for a frame uses 120 milliseconds of future audio, while the turn-ending score uses no lookahead. An application can watch for an ending without waiting for that retrospective speech score.
That second signal is useful in the unfinished sentence about the train. The final word and the speaker’s delivery both carry information that a silence timer throws away. Yet a completed thought still leaves an editorial choice for the application: should the agent take the floor immediately, offer a brief acknowledgement or wait? SAVA responds to any voice. Deciding how to treat an interruption or an encouraging “mm-hm” remains part of the conversation design.
Sources: Inworld / GitHub [6] · Inworld / Hugging Face [7] · Inworld / X [2]
What a team can take home
The two open models ship in separate repositories. Inworld links the code on GitHub and the weights on Hugging Face, so a team can run the isolator on recorded dialogue while evaluating SAVA independently on real conversational pauses. The 48 kHz output from the first model and the 16 kHz input to the second are another reason to pay attention to the handoff if a team chooses to connect them.
A third item in the announcement concerns the voice the agent speaks with. Inworld published a technical report on Realtime TTS-2 covering delivery instructions, keeping a reference voice across languages and streaming synthesis. The downloadable weights announced that day belong to Voice Isolator and SAVA. For someone evaluating the exchange about the train, those two releases already make a useful experiment possible: keep the microphone recording fixed and change when the assistant decides to speak.
Sources: Inworld [1] · Inworld / Hugging Face [5] · Inworld / Hugging Face [7] · Inworld / GitHub [8]
SOURCES & METHOD
The record behind the article
Checked 18 November 2026. Technical claims are grounded in developer documentation; videos are credited to their creators. Company claims and third-party reviews are attributed in the text; conclusions are editorial analysis. No original comparative testing of these releases was conducted for this article.
- Open-source voice models from Inworld Inworld
- November 18 release and original video Inworld / X · 2026-11-18
- Voice Isolator — two model variants and listening demo Inworld / GitHub
- Voice Isolator example 14 — original WAV input Inworld / GitHub
- Voice Isolator model card and weights Inworld / Hugging Face
- SAVA — streaming speech and end-of-turn detection Inworld / GitHub
- SAVA model card and weights Inworld / Hugging Face
- Realtime TTS-2 technical report Inworld / GitHub


