Changelog

11.1.7 adds Amazon Nova Sonic as a speech-to-speech model, three text-to-speech vendors (Fish Audio, Giggy and Svara) and KugelAudio Clarity noise isolation. Agents can now use early generation together with tools, can end the conversation after a number of unanswered prompts, and can recover from a barge-in that turns out to be noise. Tool calls from agents and Dialogflow get a longer, single timeout. The release also masks more secrets in debug logs, closes a cross-account app-{sid} call path, and fixes a scale-in drain that could wait for the full lifecycle hook.

New Features & Improvements

1. Amazon Nova Sonic — the llm verb accepts vendor: 'nova' for Amazon Nova 2 Sonic and Nova 2.5 Sonic on Amazon Bedrock, with your own AWS credentials in auth. It supports tools (asynchronous, so the agent keeps talking while your toolHook runs), polyglot voices, keypad input through llmOptions.dtmf, and calls longer than Bedrock’s 8-minute stream limit.
2. Fish Audio text-to-speech — fishaudio is available as a TTS vendor, with streaming TTS and speech credentials in the portal and API. A voice is a cloned-voice id (reference_id); the portal lists your account’s own voices. See Fish Audio.
3. Giggy text-to-speech — giggy is available as a TTS vendor, with streaming TTS and a curated voice list covering more than 80 languages. See Giggy.
4. Svara text-to-speech — svara (Kenpath Labs) is available as a TTS vendor, with streaming TTS for Hindi and other Indic languages. See Svara.
5. KugelAudio Clarity noise isolation — kugelaudio is a third noiseIsolation vendor, alongside krisp and rnnoise, in the agent and config verbs. It uses a KugelAudio speech credential with Use for noise isolation ticked, or a key passed on the verb in auth. See KugelAudio Clarity.
6. Early generation with tools — agents that use tools now get early generation (earlyGeneration: true). A speculative reply can choose tool calls, but they reach your toolHook only once the caller’s turn is confirmed, and each runs at most once. See Early Generation.
7. Agent noResponseMaxPrompts — ends the agent after a set number of no-response reprompts that the caller never answers.
8. Agent barge-in recovery (opt-in) — set JAMBONES_AGENT_BARGEIN_RECOVERY_MS (for example 2000) to recover when an interrupt-prediction barge-in fires on noise and the caller then says nothing: the agent picks up where it left off or checks the caller is still there, instead of leaving the line silent. Each recovery is counted as barge_in_recoveries in the session summary.
9. One tool-call timeout — JAMBONES_TOOL_CALL_TIMEOUT_MS (default 20000) bounds every agent and Dialogflow (CES and CX) tool call, over websocket or HTTP. Dialogflow tool calls previously timed out after 5 seconds over websocket and 10 seconds over HTTP. The new timeout only ever lengthens the existing ones.
10. Hold and unhold events (opt-in) — set JAMBONES_HOLD_UNHOLD_EVENTS to detect in-dialog hold and unhold re-INVITEs. When sipRequestWithinDialogHook is configured, the application receives a hold or unhold event.
11. Noise isolation alert types — a missing KugelAudio credential is reported as a no-noise-isolation alert, and a noise-isolation vendor that fails to start as noise-isolation-failure, instead of as a Krisp alert. Krisp keeps no-krisp.
12. Model lists — the KugelAudio TTS model list adds kugel-3.5, and Cartesia’s retired models are replaced with sonic-3.6.

Bug Fixes

• Security — a SIP client calling app-{sid} can now reach only applications in its own account, or service-provider applications in its own service provider. Debug logs now mask more secrets: Dialogflow service-account credentials in verb payloads, listen and conference websocket passwords, and recorder passwords.
• Scale-in drain — a call that hung up while an agent’s barge-in gather was running could stay counted as active, so the drain waited for the full lifecycle hook. The session is now always released.
• Answering machine detection — English voicemail fragments and read-out phone numbers are recognized, and voicemail phrases in every language are matched when the recognizer language is multi.
• Tracing — JAMBONES_OTEL_ENABLED=0 or false turned tracing on. Only 1 or true now enables it.
• SBC inbound — a call from a carrier that uses registration is matched to that carrier even when another account’s static inbound gateway range covers the same IP address. On Five9 Voicestream calls, only the audio sent back to Five9 is blocked; the whole call was being muted, so the listen websocket received no audio.
• drachtio — fixes for crashes on tel: URIs, a 491 race when retrying with authentication, a 500 during call setup, and DNS queries left orphaned on failover; general SIP stack hardening; TLS certificates reload on SIGHUP; and a validated license is kept through Redis or MySQL outages.
• Portal sign-in — OAuth sign-in is more reliable and no longer counts against the sign-up rate limit.
• Recordings — failed uploads to S3-compatible storage now log the HTTP status, so a 403 from a proxy or firewall in front of the bucket is easy to spot.

Upgrade Notes (self-hosted)

• Database — 11.1.7 adds two columns to speech_credentials (use_for_noise_isolation and noise_isolation_tested_ok); run the database upgrade before starting the new API server.
• FreeSWITCH — the feature server no longer supports FreeSWITCH. JAMBONES_FREESWITCH is ignored (a warning is logged), and the feature server connects to mediajam at JAMBONES_MEDIAJAM, default 127.0.0.1:8021.
• Tracing — installs that set JAMBONES_OTEL_ENABLED=0 had tracing on; after upgrading it is off. Set it to 1 if you want tracing.
• Tool calls — Dialogflow tool calls now wait up to 20 seconds by default (JAMBONES_TOOL_CALL_TIMEOUT_MS).

Component Versions

Availability

jambonz.cloud - coming soon

self-hosted on AWS - coming soon

self-hosted on Azure - coming soon

self-hosted Debian packages - coming soon