VSIP

Changelog

What shipped.

Four products on one engine, in the order they arrived.

2026-07

Batch Audio Analytics

  • New: POST /v1/analyze — upload a recording, get a structured conversation timeline back.
  • Timelines now carry a summary block: turn count, per-speaker talk-time ratio, total silence, longest monologue.
  • Jobs and results persist and are retained for 30 days.
  • Usage metered by audio-minute, with GET /v1/analyze/usage to check your allowance.
  • AnalyzeClient added to both the Python and TypeScript SDKs.

2026-07

Speaker Profile API

  • New: enroll, verify (1:1) and identify (1:N) against persistent voiceprints.
  • Quality gates reject audio that is too short, too quiet, or not speech — before a profile is created.
  • Consent flags and per-profile retention windows, with an automatic purge.
  • Per-request match threshold for tuning false accepts against false rejects.
  • Streams can be seeded with ?profile_id= to lock identity from the first word.
  • ProfilesClient added to both SDKs.

2026-06

SDKs

  • Released vsip-sdk on PyPI (import vsip) and vsip-sdk on npm.
  • Adapters: Pipecat (VSIPProcessor), LiveKit (VSIPVAD), Twilio Media Streams (TwilioBridge).
  • Clients handle reconnect and resume, lookback slicing, speak-gating and turn queueing.

2026-05

Voice Events API

  • Identity-gated barge-in — only the verified speaker can interrupt the agent.
  • Semantic endpointing: turns end on meaning, not on a silence timer.
  • Noise classification keeps TV, music and keyboard sounds out of the LLM.
  • sidecar mode for listen-only use beside speech-to-speech stacks.
  • Server-side echo cancellation (aec=server) for speakerphones and kiosks.