Skip to content

v0.2.0

Choose a tag to compare

@github-actions github-actions released this 04 Nov 11:51
· 97 commits to main since this release
cc3afff

What changed for you

Faster first audio in the LiveKit plugin. speko_gateway.livekit.TTS now keeps one Gateway session open and ready for the next reply. A reply's first text no longer waits for a session and its provider connection to open. With managed ElevenLabs, that also takes the provider's token check out of the path to first audio. In direct measurements of that connection from Southeast Asia, first audio on eleven_v4_turbo took 553 ms P50 when text was sent right after connecting, and 59 ms on a connection that was already open.

What you need to do. Rebuild your agent image on spekoai/gateway:0.2. No code changes are needed. LiveKit calls prewarm() when your agent starts, and that opens the first standby session.

One more open session per agent.

  • Between replies, each TTS holds one extra Gateway session and provider connection. During a reply it holds two.
  • These count toward your organization's concurrent-session limit, the Gateway's SPEKO_MAX_SESSIONS, and provider connection limits.
  • With managed credentials, each open session also reserves credit for its max_input_characters.
  • If you run close to any of these limits, raise them before you upgrade, or lower max_input_characters.

Opting out. TTS(standby_sessions=0) restores the previous behavior.

Cleanup. A standby that no reply uses closes after 3 minutes. LiveKit does not close the TTS when a session ends, so call aclose() to release it sooner.

The Pipecat integration is unchanged in this release.

Image: spekoai/gateway:0.2.0 (also :0.2)

What's Changed

  • Versioned releases with a changelog customers can watch by @Baymurat785 in #145
  • feat(livekit): keep a standby TTS gateway session for the next reply by @Laroikin in #146

Full Changelog: v0.1.0...v0.2.0