New Features & Improvements
- New speech vendors — speech-to-text:
sarvam,smallestai,gnani,meta,saidwire,azureopenaiandinworld. Text-to-speech:sarvam,smallestai,gnani,minimax,neuphonic,asyncai,respeecher,camb,hume,blandandsonioxwith streaming TTS, andvakyam,simplismart,groq,speechmatics,azureopenai,kokoroandspitchforsay. Each has speech credentials in the portal and API. See TTS vendor settings. - KugelAudio speech-to-text —
kugelaudio, already a text-to-speech and noise isolation vendor, now provides streaming speech-to-text (kugel-ear-1). See KugelAudio. - Soniox speech-to-text v6 — Soniox credentials and recognizers can use the
stt-rt-v6model. - Speech-to-speech vendors in Agent Studio — Agent Studio adds Google Gemini (with a Gemini API key or a Vertex AI service account), xAI, Qwen, AssemblyAI, Azure VoiceLive, Azure OpenAI and Amazon Nova as realtime voice agents. These vendors were already available in the llm verb; each now has an account credential in the portal (LLM Services, Speech-to-speech tab) and a key test that does not open a billable session. These agents report turns to Agent Studio like the existing ones. See Realtime voice API.
- Sign in with GitHub or Google — a portal user can be created or edited to sign in with
GitHub or Google instead of a password. The user receives an invitation email and is
linked to their account on first sign-in by verified email address. Switching a user to
GitHub or Google removes their password; switching back requires a new initial password,
which they must change at first sign-in.
GET /Usersreports each user’sproviderand whether the invitation is still pending. - OpenAI Realtime errors — an OpenAI Realtime agent now reports an
engine_errorturn event when the vendor returns an error, a response fails, or the connection fails or drops, and the verb’s actionHook carries anerrorfield. - Speech-to-speech close reason — when a speech-to-speech vendor closes the connection, the disconnect now includes the vendor’s close code and reason.
- API version —
GET /v1/Versionreturns the api-server release. - Portal — call traces mark Dialogflow CES turns that the agent started without caller speech, and the hosted portal has the service provider switcher back for admins.
Changes in Behavior
- Turn detection by the speech vendor — with the agent verb, speech-to-text
vendors that do their own turn detection (
deepgramflux,assemblyai,speechmaticsagent,zoom,gradium,mistral,paxaand others) now end the caller’s turn on the vendor’s end-of-turn signal only, not also on the final transcript, which could run the turn twice. An end of turn with no words is ignored. - Azure OpenAI realtime endpoint — Azure OpenAI realtime agents connect to the GA
endpoint (
openai/v1/realtime?model=) instead of the previewapi-versionanddeploymentendpoint. - Gemini tool calls — Gemini realtime agents send one toolHook per tool call, carrying the vendor’s call id, instead of one per batch.
- Tenant checks — changing a SIP gateway’s carrier with
PUT /SipGatewaysnow returns 400. Assigning a phone number to an account outside your scope returns 403, and assigning it to an unknown account or application, or to an application of a different account, returns 400. - Credential validation — a speech credential request missing a required field returns 400 instead of 500. LLM credentials that have never been tested now show as untested rather than failed, and are tested again when the portal lists them.
Bug Fixes
- Hold on anchored-media dials — when a dial anchors media, the caller’s hold
and resume are now answered by jambonz and passed to the
onHoldHook. - Azure speech-to-text socket leak — speech-to-text vendor connections are now closed when recognition stops. Azure connections were left open, which over time caused gaps of up to 1.3 seconds in text-to-speech.
- ElevenLabs data residency — ElevenLabs speech-to-text, and its credential test, use the host set by the credential’s data residency.
- Deepgram Flux — Flux uses the recognizer’s model and language.
- Speech-to-text connection failures — alerts and errors now give the vendor’s reason.
- AssemblyAI agents — an AssemblyAI agent can be asked to reply mid-call.
- VAD barge-in — when VAD barge-in stopped the agent for a sound the speech-to-text vendor never transcribed, the agent now resumes instead of staying silent.
- OpenAI speech-to-text — OpenAI’s speech-stopped event now reaches the agent.
- Dialogflow CES traces — traces of long CES calls are no longer dropped.
- License changes — calls are no longer refused while a license change is being applied.
- Recent calls — the list no longer empties after the date filter changes.
Availability
jambonz.cloud - November 7, 2026
self-hosted Debian packages - available now: install jambonz-mini 11.1.12 from the
jambonz apt repository (Debian 12 bookworm), amd64 and arm64.
self-hosted on AWS - available now: CloudFormation templates and public AMIs for all nine server types in 30 regions, amd64 and arm64.
New Features & Improvements
- Saidwire speech-to-text —
saidwireis available as a realtime STT vendor, with speech credentials in the portal and API. Saidwire follows Deepgram’s streaming protocol and is handled likedeepgram: interim results, a final transcript per utterance, and jambonz’s own turn detection in the agent verb. It is English only and mono, and reports a fixed confidence of 0.95.saidwireOptionssetspunctuate,smartFormatting,endpointing,utteranceEndMsandvadEvents. See recognizer.
New Features & Improvements
- Sarvam AI text-to-speech —
sarvamis available as a TTS vendor, with streaming TTS and speech credentials in the portal and API. The Bulbul modelsbulbul:v4-flash(default),bulbul:v3andbulbul:v2cover eleven Indian languages, and the portal lists thebulbul:v4-flashpersonas by language. Options arepace,pitch,loudness,temperature,enable_preprocessinganddict_id, plus buffering controls for streaming. See Sarvam AI. - Sarvam AI speech-to-text — the same
sarvamcredential serves realtime speech recognition withsaaras:v3-realtime(default) orsaaras:v4, in 23 Indian languages or with automatic language detection. Sarvam signals the start and end of speech and returns a final transcript after each pause.sarvamOptionssets the stream type, the transcript mode (including translation to English), the voice activity settings, a prompt and, forsaaras:v4, key terms. See recognizer.
New Features & Improvements
- Smallest AI text-to-speech —
smallestaiis available as a TTS vendor, with streaming TTS and speech credentials in the portal and API. The Lightning modelslightning_v3.1_pro(default),lightning_v3.1_pro_10_26andlightning_v3.1are supported, and the portal lists the Pro voices by language. Options arespeed,language,number_pronunciation_language,date_format,math_notation,pronunciation_dictsandcontent_filter, plus buffering controls for streaming, and word-level timestamps make alignment events available for the voices that return them. See Smallest AI. - Smallest AI speech-to-text — the same
smallestaicredential serves realtime speech recognition with Pulse (pulse, default, or the English-onlypulse-2).hintsare sent as Pulse’s keywords (up to 100 terms), andsmallestaiOptionssets formatting, end-of-utterance timing, diarization, redaction, timestamps and detection features. See recognizer.
New Features & Improvements
- Azure OpenAI text-to-speech —
azureopenaiis available as a TTS vendor for one-shotsay(no TTS streaming), with speech credentials in the portal and API. A credential holds the resource key, the resource endpoint and the deployment names. The default deployment isgpt-4o-mini-ttswith the voicealloy, and options arespeed,instructionsandmodel_id. See Azure OpenAI. - Azure OpenAI speech-to-text — the same
azureopenaicredential serves realtime speech recognition on a transcription deployment (gpt-live-transcribeby default, orgpt-realtime-whisper,gpt-4o-transcribe,gpt-4o-mini-transcribe).azureopenaiOptionsmirrorsopenaiOptions: deployment, model family, prompt, languages, keywords, delay, noise reduction, turn detection and local voice activity settings. See recognizer.
New Features & Improvements
- Gnani text-to-speech —
gnaniis available as a TTS vendor for one-shotsay(no TTS streaming), with speech credentials in the portal and API. The Timbre modelstimbre-v2.5(default) andtimbre-v3.0cover ten Indian languages plus Hinglish, and the portal lists the voices by language. Options aremodel_id,languageandspeed. See Gnani. - Gnani speech-to-text — the same
gnanicredential serves realtime speech recognition withprisma-v2.5(default) orprisma-v3.0, in ten Indian languages or with automatic Indic detection. Gnani signals the start and end of speech and returns one final transcript per utterance.hintsare sent as bias words or key terms, andgnaniOptionssets the transcript format, substitutions, context, idle timeout and voice activity settings. See recognizer.
New Features & Improvements
- Meta speech-to-text —
metais available as an STT vendor, with speech credentials in the portal and API, for realtime speech recognition with Meta’smuse-voice-transcribe-1.0. Meta signals the start of speech, sends interim transcripts and returns one final transcript per turn. The recognizer’slanguageis sent as a language bias,hintsare sent as keywords, andmetaOptionssets the mode, the partial transcript style, keywords, language bias and a session id. See recognizer.
New Features & Improvements
- MiniMax text-to-speech —
minimaxis available as a TTS vendor, with streaming TTS and speech credentials in the portal and API. The Speech modelsspeech-2.8-turbo(default) throughspeech-01-hdare available, and the portal lists MiniMax’s system voices by language. Options include speed, volume, pitch, emotion, text normalization, language boost, voice effects, pronunciation dictionaries and voice mixing. Streaming sends text over a live connection and cancels on barge-in. See MiniMax.
New Features & Improvements
- Neuphonic text-to-speech —
neuphonicis available as a TTS vendor, with streaming TTS and speech credentials in the portal and API. The portal lists Neuphonic’s stock voices by language, and account voices can be used by their voice id. Options arelanguage,speed,temperature,ipa_keyand, for streaming,timestamps. Streaming returns word-level timestamps, so alignment events are available. See Neuphonic.
New Features & Improvements
- Camb.ai text-to-speech —
cambis available as a TTS vendor, with streaming TTS and speech credentials in the portal and API. Models aremars-flash(default),mars-pro,mars-instructand themars-8.1betas; voices are numeric ids, and cloned voices work by their id. Options includespeaking_rate, enhancement switches,inference_stepsand themars-instructuser_instructions. Streaming returns word-level timestamps, so alignment events are available. See Camb.ai.
New Features & Improvements
- Async text-to-speech —
asyncaiis available as a TTS vendor, with streaming TTS and speech credentials in the portal and API. Models areasync_flash_v1.5(default),async_flash_v1.0andasync_pro_v1.0; voices are Async voice ids. Options aremodel_id,language,speed_controlandstability. Streaming does not return word timestamps, so alignment events are not available. See Async.
New Features & Improvements
- Respeecher text-to-speech —
respeecheris available as a TTS vendor, with streaming TTS and speech credentials in the portal and API. Models areen-rt(default),ua-rt,ar-rt,de-rt,pt-br-rtandsv-rt, picked from the language unlessmodel_idis set. Options aremodel_idand the sampling controlsseed,temperature,top_k,top_p,min_p,presence_penalty,repetition_penaltyandfrequency_penalty. Streaming does not return word timestamps, so alignment events are not available. See Respeecher.
New Features & Improvements
- Bland AI text-to-speech —
blandis available as a TTS vendor, with streaming TTS and speech credentials in the portal and API. The voice is a Bland voice UUID, which also selects the model; Karen is the default en-US voice. Options areexpressivenessandstability. Streaming does not return word timestamps, so alignment events are not available. See Bland AI.
New Features & Improvements
- Simplismart text-to-speech —
simplismartis available as a TTS vendor for one-shotsay(no TTS streaming), with speech credentials in the portal and API. Models are Orpheus 3B (default), Veena for Hindi and Qwen 3 TTS on a dedicated deployment, set with the credential’s API host. Options aremodel_id,temperature,top_p,repetition_penaltyandmax_tokens, pluslanguageandleading_silencefor Qwen. See Simplismart.
New Features & Improvements
- Spitch text-to-speech —
spitchis available as a TTS vendor for one-shotsay(no TTS streaming), with speech credentials in the portal and API. Voices cover English, Yoruba, Hausa, Igbo, Amharic and Nigerian Pidgin;linais the default en-US voice. Options arespeedandapi_uri. See Spitch.