11.1.16 covers 11.1.13 through 11.1.16. It adds a preview of jambonz’s own turn detection and
interrupt prediction models for the agent verb as an alternative to Krisp, ai-coustics noise isolation and
VAD, Phonic for text-to-speech and speech-to-speech, KugelAudio Fluid speech-to-speech, and
ElevenLabs Scribe v3 Live with predicted turn ends. Agent applications receive more events: whether
each agent:update was applied, turn-taking signals, and the end of turns the agent started
itself. The release fixes two API authorization issues, a gather inter-digit timeout bug, and a
database upgrade from 10.0.x that never completed.
New Features & Improvements
- jambonz turn detection (preview) — the agent verb accepts
turnDetection: "jambonz"(or{mode: "jambonz", threshold}), which runs jambonz’s own end-of-turn model in the media server instead of Krisp. The default threshold is 0.9 and falls over 1.6 seconds of silence. This is a preview: the model and its defaults may change in later releases. - jambonz interrupt prediction (preview) —
bargeIn: {strategy: "interruptPrediction", vendor: "jambonz"}runs jambonz’s interrupt model, also a preview. The interrupt vendor must match the turn detection vendor; a mismatch falls back to VAD barge-in with a warning. - ai-coustics noise isolation and VAD —
aicousticsis a noise isolation vendor (noiseIsolationon the agent and config verbs) and a VAD vendor, with native 8 kHz and 16 kHz models. Self-hosted systems need an ai-coustics license key. - Phonic —
phonicis a text-to-speech vendor forsay(no streaming) and a speech-to-speech vendor for the llm verb, withphonicspeech credentials andphonic_s2sLLM credentials in the portal and API. See Phonic. - KugelAudio Fluid speech-to-speech —
kugelaudiois a speech-to-speech vendor for the llm verb. Withoutauthon the verb, the account’skugelaudio_s2scredential is used, and its region (global,eu,us) picks the host. - ElevenLabs Scribe v3 Live —
scribe_v3_liveis available as an ElevenLabs speech-to-text model. With the agent verb it defaults to theturn_predictioncommit strategy and ends the caller’s turn on each commit, about 0.8 seconds sooner than VAD on a complete question. - Agent update results — every
agent:updatenow sends anagent_updateevent to the eventHook saying whether it was applied, and why not when it was rejected. - Turn-taking events — the agent eventHook now receives
overlapping_speech,user_interruption(with its cause),interrupt_prediction_ignored,no_response_timeoutandllm_error, which were previously only in feature-server logs. - Agent-started turns — a turn the agent started itself (the greeting, or a reply it was
asked to generate) now sends
turn_endwith its response andagent_initiated: true. - Billing — account users on a paid plan can update their card from the Billing overview (jambonz.cloud).
Changes in Behavior
- OpenAI Realtime default model — an llm verb with
vendor: openaiand nomodelnow usesgpt-realtime-2.1. OpenAI retiresgpt-realtimeon 2027-01-20.vendor: microsoftkeepsgpt-realtime, since there the model is the Azure deployment name. - Gather inter-digit timeout — in gather, the first key now stops the
no-input timeout and starts
interDigitTimeout.minDigits(elsenumDigits) only decides whether the entry counts; below it the result is a timeout, and the timeout actionHook carries the digits entered.interDigitTimeoutnow works withoutminDigits. - Credential labels — a speech or LLM credential label needs to be unique only within an account and vendor; a label on a sibling account no longer blocks it.
- Free trials — on jambonz.cloud one free trial is allowed per mailbox:
+tagaliases, and dots in Gmail addresses, count as the same address.
Security Fixes
- API authorization —
/v1/logingave some account users a wider scope than their account. It now derives scope the same way as/signinand API keys. - View-only users — a view-only user who signed in to the portal with email and password could make changes. View-only is now enforced for every sign-in method.
- Query parameters — object-valued query parameters and request fields are now treated as plain strings in database queries, so they can no longer widen a query across tenants. Self-hosted 500 responses no longer include internal error details.
Bug Fixes
- Database upgrade from 10.0.x — upgrading a 10.0.x database to 11.1.x never completed
because the
krisp_usageandllm_credentialstables were skipped. They are now created whenever they are missing. - Agent tools updates —
update_toolsno longer removes the agent’s handoff and hangup tools. - Acoustic turn detection — with Krisp or jambonz turn detection and a speech-to-text vendor that sends no speech start (Deepgram), the LLM sometimes received only part of what the caller said. The whole turn is now kept.
- Speech-to-text restarts — a transcription that the vendor closed (Google’s 305-second limit) is restarted; a restart that fails is now reported to the application as an ASR error instead of stopping silently.
- Answering machine detection — three voicemail greeting phrases that were detected as a
human (
record your message,telephone number and your message,número de teléfono) are now built-in AMD hints. - OAuth sign-in — a GitHub or Google sign-in that matches no user now shows why on the login page, naming the verified address, instead of sending the person to sign up for a trial. An invited user is no longer signed in to a trial of their own.
- Grafana dashboards — on AWS mini, medium and large images since 11.1.6, every panel of the
Jambonz Metrics dashboard showed
Datasource ... was not found. The dashboards now use the provisioned InfluxDB datasource. - Websocket reconnects — an unanswered
session:reconnectno longer produces unhandled rejection log errors. - SIP — a pending NOTIFY is no longer answered with 500 when a BYE ends the dialog, TCP 2xx responses are retransmitted until the ACK arrives, and stale TLS errors no longer affect new connections.
Upgrading
There are no database schema changes since 11.1.12.
Component Versions
Availability
self-hosted on AWS - available now: CloudFormation templates and public AMIs for all nine server types in 30 regions, amd64 and arm64.
self-hosted on Azure - available now: Terraform modules and Compute Gallery images for all nine server types, amd64 in 30 regions and arm64 in 29.
self-hosted on GCP - available now: Terraform modules and public images for all nine server types, amd64 and arm64. GCP images are global, so there is no per-region list.
The Azure and GCP images ship mediajam 0.5.23, which fixes a crash at startup when the optional jambonz turn-taking models are not installed.