11.1.6 adds alert notifications, so alerts reach you as they happen, and lets
customers opt out of support audio capture. It adds Speechify and KugelAudio
text-to-speech, Zoom and Gradium speech-to-text, ElevenLabs eleven_v3/eleven_v4 streaming
and AssemblyAI Universal-3.6 Pro. The API server now rate-limits per account and
per caller. Calls through a carrier that requires registration now leave from the SBC
that holds the registration. On AWS, scale-in drains SIP, RTP and feature servers by reading the
instance metadata service instead of subscribing to SNS. The release also fixes
a set of agent, answering machine detection and
gather issues.
New Features & Improvements
- Alert notifications — every alert on an account’s Alerts page can also be
pushed, as it happens, to up to five destinations: a JSON webhook, Slack,
PagerDuty or email. Add them in the account’s Settings → Notifications, and
send a test to any one destination. An account is notified about the same
type of alert at most once an hour. The same destinations can be managed
through the API (
/Accounts/{AccountSid}/AlertHooks). - Support audio capture opt-out — jambonz support keeps SIP and RTP from calls for a few days to troubleshoot call quality. An account can now turn this off in Settings → Privacy, and an organization can allow it for all accounts, disable it for all accounts, or leave it to each account. Recent Calls SIP traces are not affected.
- Speechify text-to-speech —
speechifyis available as a TTS vendor, with speech credentials in the portal and API. - KugelAudio text-to-speech —
kugelaudiois available as a TTS vendor, including streaming TTS. - Zoom Scribe speech-to-text — Zoom is available as an STT vendor.
- Gradium speech-to-text —
gradiumis available as an STT vendor, with speech credentials in the portal and API. - ElevenLabs
eleven_v3andeleven_v4— streaming TTS with these models, includingeleven_v3_conversationaland its audio tags (such as[laughs]), uses ElevenLabs’ Text to Dialogue stream. The synthesizer language is now sent to ElevenLabs instead of being detected on each turn, and the models are listed in the portal. - AssemblyAI Universal-3.6 Pro — new AssemblyAI credentials default to
universal-3-6-pro, andassemblyAiOptionsacceptslanguageCodesandvoiceFocus. Newer models are now sent 16 kHz audio instead of 8 kHz. - ElevenLabs agent keypad input — the ElevenLabs llm verb
accepts
llmOptions.dtmf. Withenable: true, the caller’s key presses are collected into one entry, using the gather digit options (finishOnKey,numDigits,maxDigits,interDigitTimeout,timeout), and sent to the agent as a single user message. Mid-callllm:updatemessages are now passed to ElevenLabs. - Dialogflow CES session variables — for
model: "ces", the dialogflow verb sendswelcomeEventParamsto the agent as session variables when the session starts, and no longer requires awelcomeEvent. - OpenAI GPT-Live — the GPT-Live speech-to-speech integration uses OpenAI’s GA protocol.
- Agent no-response limit — the agent verb’s new
noResponseMaxPromptsends the agent after that many no-response prompts in a row get no reply, and itsactionHookreceivescompletion_reason: "no_response". Unset, the agent keeps prompting as before. - Per-account API rate limits — the API server counts requests per account
for call creation (4000/min) and call status reads (600/min), per caller
address for everything else, and per address for sign-in and sign-up. Counts
are shared by all API server workers, and a client over a limit receives
429. Every limit can be changed with an environment variable. - SRTP for TLS-registered devices — calls to a SIP client that registered
over TLS are offered SDES-SRTP, so encrypted signaling gets encrypted media.
Set
JAMBONES_DISABLE_SRTP_FOR_TLSto keep plain RTP. (jambonz.cloud keeps plain RTP for now.) - AWS scale-in without SNS — SIP servers, RTP servers and feature servers drain on scale-in by polling the instance metadata service for their target lifecycle state. They no longer subscribe to an SNS topic or open a port for it. When a SIP server holding the carrier registrations scales in, it hands them to another SIP server before it stops. Registrations now use a Call-ID per SIP server, so a carrier sees a move between servers as a new registration.
- Registration trunks use the registering SBC — when a call goes out through a carrier that requires registration, jambonz now sends the INVITE from the SBC that holds the registration, so carriers that accept calls only from the registered address (BroadWorks, for example) no longer reject it. This applies to dial, calls created through the API, and the SBC’s own carrier selection.
- Agent end-of-turn fallback (opt-in) — set
JAMBONES_AGENT_EOT_FALLBACK_MS(for example3000) to take the caller’s turn on the final transcript when a turn detector such as Krisp never reports the end of the turn. Each fallback is logged and counted aseot_fallbacksin the session summary. - Enterprise user limits — an organization’s user limit can be raised per organization, and the portal’s Users view follows the limit the API reports.
- SIPREC application in the new portal — the account setting for the application that handles SIPREC calls is back in the new portal.
- Self-hosted updater — the update client upgrades itself when a release requires it and can apply several releases in one run.
- Security hardening — service-provider users can no longer write operator-only organization or account fields; carrier changes by non-admin users stay within their own organization or account; and calls from a registration trunk’s address now match only carriers that belong to the called SIP realm’s account. New installs generate a random admin API token instead of seeding a fixed one.
- Fewer caller details in logs — caller transcripts and DTMF digits are
logged at
debug, notinfo, and the credentials for a listen websocket are no longer written to debug logs.
Bug Fixes
- Agent verb — the caller’s turn ends normally when Krisp is used only to predict interruptions; speech that arrives while the assistant is talking is held for the next turn instead of being dropped; interrupt prediction only interrupts the assistant while it is speaking; and VAD is used only for turn timing when the STT vendor reports the start of speech itself. With parallel tool calls, the LLM is prompted once when every result is in, instead of once per result. What the caller said before a pause for more speech is kept, speech held over the assistant is kept once, and a turn that was already answered is not sent to the LLM again.
- Answering machine detection — empty final transcripts no longer count as a human, Deepgram endpointing defaults to 150 ms during detection, and built-in voicemail hints are used when none are configured.
- Gather with Deepgram — an interim word that Deepgram later drops no longer
delays every following
UtteranceEnd, and a deferredUtteranceEndreturns the transcript as soon as it is satisfied. - Soniox — v5 sub-word tokens are joined correctly (“I have Optum.”, not “I have O pt um.”).
- Google Gemini transcription —
SMARTmode no longer fails withlanguageCodesset. - Deepgram warm connections — options that change between transcriptions (for example new Flux keyterms) are applied when a parked connection is reused.
- TTS — a say no longer hangs when the TTS vendor rejects
the request, and streaming TTS starts cleanly after a barge-in. Streaming TTS events and
TTS_STREAMING_CONNECTION_FAILUREalerts name the vendor actually in use. - Conference —
distributeDtmfandspeakOnlyTowork without another join option set. - Dial — when two forked calls answer at the same time, the second is released
instead of leaking or replacing the first. A dial with
exitMediaPath: truefalls back to a partial media release whenJAMBONES_ENABLE_FULL_MEDIA_RELEASEis not set. - Feature server shutdown — on
SIGTERMorSIGINT, a feature server drains calls in progress instead of exiting at once. - SBC resource leaks — a DTMF socket leak in the SBC (
@jambonz/rtpengine-utils0.4.5) that could exhaust file descriptors and cause500responses is fixed. An outbound call to a carrier with no outbound gateways no longer leaves an rtpengine session behind, and rtpengine commands time out after 5 seconds, so a lost reply cannot hang a call or hold its place in the account’s call limit. - Session tokens — on deployments with separate SBC and feature servers, a small clock difference between the hosts could leave an outbound call answered with no audio. Session tokens now allow up to two seconds of clock skew (drachtio 10.1.8, mediajam 0.5.14).
- Media timeout —
JAMBONES_MEDIA_TIMEOUT_MSnow ends calls on mediajam feature servers. - Dialogflow — a dialogflow verb with no TTS vendor no longer looks up TTS credentials.
- Debug logging — every leg of a call follows the account’s debug log setting.
- SIP — a re-INVITE rejected with
422no longer leaks resources in drachtio; licensed SBCs issue a fresh session token with each provisional and final response, so calls that ring longer than 40 seconds keep their audio; TLS registrations record their transport; and calls forwarded to asips:Request-URI use asips:Contact and From. - Billing — after a failed renewal the portal shows the suspended subscription and asks for a new card, instead of offering an upgrade that created a second subscription.
- Recent Calls pcap download — a download the client abandons no longer leaves connections to HOMER open, and requests to HOMER time out after 30 seconds.
- Forced password change — on jambonz.cloud, a user created with a temporary password is now asked to set a new one at first sign-in. The Create password page asks for the temporary password again instead of keeping it in browser storage.
- Portal — the call logs panel tells a failed request apart from a call with no logs.
- License keys — whitespace pasted with a license key is trimmed.
Self-hosted on AWS (CloudFormation)
The 11.1.6 CloudFormation templates are designed to enable SOC 2 compliance for self-hosted deployments on AWS. They close SSH by default, take every secret out of files and user data, give each server only the AWS permissions it needs, encrypt and protect the database, keep logs for a set period, and keep the SBCs’ published addresses stable. They also add a load-balanced web tier for large deployments.
Key changes
- SSH closed by default —
AllowedSshCidrhas no default; supply your IP as a/32or your VPN range. - Secrets in SSM Parameter Store (medium, large) — every secret is kept in SSM
SecureString parameters under
/jambonz/<stack>/; none is written into an ecosystem file, a systemd unit or user data. The database password is generated into Secrets Manager. - Least-privilege instance roles (medium, large) — each tier’s role grants only
the AWS calls its software makes;
AmazonEC2FullAccessandAutoScalingReadOnlyAccessare no longer attached. - Database protection — Aurora is encrypted on both sizes (large was unencrypted), backups are kept for 30 days, deletion protection is on, and a final snapshot is taken when the stack is deleted. Large logs database connections (logins, logouts, failed logins) to CloudWatch.
- Log retention enforced —
CloudwatchLogRetentionnow takes effect (the instance roles lacked the permission to apply it). The default is 90 days, and 180 and 365 are allowed. - Multi-AZ Redis (large) — Redis runs as two nodes in separate availability zones with automatic failover.
- SBCs keep their published address (medium, large) — a SIP or RTP server waits for its Elastic IP and marks itself unhealthy if it cannot take it, instead of serving from a temporary public IP that carriers have not allow-listed.
- No shared credentials in the templates — the templates no longer obtain an
apiban.org key with a jambonz account; set the optional
ApibanKeyparameter to use the apiban blocklist.KrispApiKeyis now hidden in the console. - SIP over TLS and WSS — store a certificate in SSM Parameter Store
(
TLS_PRIVKEY,TLS_CERT,TLS_CHAINunder/jambonz/<stack>/sbc) and every SBC configures drachtio for TLS (5061) and WSS (8443) when it boots. Without a certificate, SIP stays UDP and TCP. - Web tier behind a load balancer (large) — set
WebServerDeploymenttoautoscaling-albto run one to four web servers in an Auto Scaling group behind an Application Load Balancer that terminates TLS. SetHostedZoneIdand the stack issues the certificate and creates the portal,api.,grafana.andpublic-apps.DNS records in Route 53, or supply your own certificate withWebCertificateArn. The default,single-instance, is unchanged. - Recording servers always deployed (medium, large) — feature servers always
send recordings to the recording servers; the fallback that ran
upload_recordingson the web server is removed. - Initial admin password in the stack outputs — in every size, the portal’s
initial password is the stack’s unique ID, shown in the
Password,PortalPasswordorWebPortalPasswordoutput. You are asked to change it at first sign-in. - Grouped parameters (medium, large) — the console lists the required parameters first, then optional settings, then groups whose defaults suit most deployments, as mini already did.
- Fixes — the
api_base_urlsent to applications in webhook payloads now points at the API (<portal>/api/v1,httpsbehind the large load balancer) instead of the portal, and turningEnableOpenTelemetryoff now stops the feature servers’ tracing.
Before you deploy or update a stack
AllowedSshCidris required on every create and update.- Medium and large need
--capabilities CAPABILITY_NAMED_IAM(it works for mini too). The commandgenerate-cf.shprints now uses it. - The
MySQLPasswordparameter is removed from medium and large; the password is generated.AicApiKeyis removed from every size. - Medium and large always deploy the recording servers, and large’s Redis runs two
nodes. Both add cost.
cache.t2.smallandcache.t2.mediumare no longer accepted for large’s Redis because they do not support Multi-AZ failover. - On an existing stack,
CloudwatchLogRetentionis now applied, so log events older than the chosen period are deleted. - Aurora deletion protection is on: turn it off before deleting a stack. The Elastic IPs, the final database snapshot and large’s database audit log group are kept after the stack is deleted; remove them yourself when you no longer need them.
Self-hosted on Azure (Terraform)
The 11.1.6 Azure Terraform modules bring the same SOC 2 controls to Azure. The medium and large modules also spread every tier across availability zones, so a deployment survives the loss of one zone.
Key changes
- SSH closed by default —
allowed_ssh_cidrhas no default; supply your IP as a/32or your VPN range. - Only the ports each server needs — on medium and large, each server role has its own network security group that lists its ports and denies all other traffic from inside the virtual network (Azure otherwise allows all traffic within a virtual network). Internal ports, including the API on 3000, are no longer reachable from the internet. On mini, Homer (9080) and port 3000 are no longer open.
- No secrets in custom data — on medium and large every secret is kept in Key Vault and each VM reads the ones it needs at boot with its managed identity. On mini the JWT secret and MySQL password are generated on the VM at first boot, and the optional apiban and Krisp keys go in a Key Vault created only when one is set. Setup scripts no longer echo their commands, so no secret appears in the setup log.
- Least-privilege identity (medium, large) — the managed identity no longer holds Network Contributor or Virtual Machine Contributor; it can only read secrets.
- Availability zones (medium, large) —
zones(default["1","2","3"]) spreads SIP and RTP servers (large) or SBCs (medium) across zones, balances the feature server and recording scale sets, makes public IPs and the recording load balancer zone-redundant, and keeps a zone-redundant MySQL standby with automatic failover. Setzones = []in a region without availability zones. - Database protection (medium, large) — MySQL logs connections (logins, logouts,
failed logins) to Log Analytics and has a
CanNotDeletelock (mysql_delete_lock). Backups are kept for 30 days and storage is always encrypted. - Private Redis (medium, large) — public network access is disabled; the virtual network reaches Redis through a private endpoint.
- Centralized logs with retention — a Log Analytics workspace keeps syslog
(including SSH logins) and the jambonz application logs for
log_retention_days(90 by default; 180 and 365 are allowed), collected by the Azure Monitor Agent on every VM and scale set. - Fixes (large) — recording servers now have outbound internet access, so recordings can be uploaded to your storage; and with more than one RTP server the SBCs no longer fail every call, because they now find RTP servers through Redis.
Before you deploy or update
allowed_ssh_cidris required.- Medium and large default to two SIP and two RTP servers (large) or two SBCs
(medium), two feature servers, and a General Purpose MySQL SKU
(
GP_Standard_D2ds_v4), because zone-redundant MySQL does not support Burstable SKUs. These add cost; to keep a Burstable SKU setmysql_high_availability = false. - On an existing deployment, adding zones replaces the public IPs, VMs and scale
sets, so the SBC addresses your carriers allow-list change. Set
zones = []to keep the current layout. terraform destroyremoves the MySQL delete lock first; deleting the resource group any other way fails until you remove it. If a destroy fails withInUseSubnetCannotBeDeleted, wait a few minutes and run it again.- The modules need the
azurermprovider 4.55 or later; runterraform init -upgrade.
Self-hosted on GCP (Terraform)
The 11.1.6 GCP Terraform modules bring the same SOC 2 controls to Google Cloud. The medium and large modules also spread every tier across zones, so a deployment survives the loss of one zone.
Key changes
- SSH closed by default —
allowed_ssh_cidrhas no default; supply your IP as a/32or your VPN range. - Only the ports each server needs — on medium and large, each server role has its own firewall rule listing only the ports it serves inside the network, replacing a rule that allowed all traffic within the network. On medium the API port (3000) is no longer open to the internet. On mini, ports 3000, 3010 (Grafana), 9080 (Homer) and 16686 (Jaeger) are no longer open; the portal, API and Grafana are reached through nginx on 80/443.
- No secrets in instance metadata — on medium and large every secret is kept in Secret Manager and each VM reads the ones it needs at boot with its service account. On mini the JWT secret and MySQL password are generated on the VM at first boot, and the optional apiban and Krisp keys go in Secret Manager only when one is set. Startup scripts no longer echo their commands, so no secret appears in the serial console or the setup log.
- Least-privilege service account — on medium and large the service account no
longer holds
roles/compute.instanceAdmin.v1on the project; it can delete only the deployment’s own feature servers (when they scale in), read its secrets, and write logs and metrics. Mini gets a service account of its own with the same limits. - Zones (medium, large) —
zones(default: the region’s first three) spreads SIP and RTP servers (large) or SBCs (medium) across zones, runs the feature server and recording instance groups across all of them, keeps a Cloud SQL standby in a second zone with automatic failover, and runs Memorystore Standard Tier with a replica in another zone. - Database protection (medium, large) — Cloud SQL logs connections (logins,
logouts, failed logins) to Cloud Logging, has deletion protection on
(
mysql_deletion_protection), keeps 30 backups, and supports point-in-time recovery. - Redis authentication (medium, large) — Memorystore requires a password, kept in Secret Manager.
- Centralized logs with retention — the Ops Agent on every VM sends the systemd
journal (including SSH logins) and the jambonz application logs to a Cloud
Logging bucket of the deployment’s own, kept for
log_retention_days(90 by default; 180 and 365 are allowed). - Fix (large) — with more than one RTP server the SIP servers no longer fail every call, because they now find RTP servers through Redis.
Before you deploy or update
allowed_ssh_cidris required.zoneis replaced byzoneson medium and large.- Medium and large default to two SIP and two RTP servers (large) or two SBCs
(medium), two feature servers, a regional (high-availability) Cloud SQL instance
and Standard Tier Memorystore. These add cost; set
mysql_high_availability = falseorredis_tier = "BASIC"to keep a single zone for either. - On an existing deployment, spreading across zones replaces the VMs and instance
groups, and moving Redis to Standard Tier replaces the Redis instance. Review
terraform planbefore you apply. - Cloud SQL deletion protection is on: set
mysql_deletion_protection = falseand apply beforeterraform destroy. - Medium and large turn on Data Access audit logs for Cloud SQL in the project, so the connection log reaches Cloud Logging.
Upgrade Notes (self-hosted)
- Database — 11.1.6 adds the
alert_hookstable and three columns; run the database upgrade before starting the new API server. - AWS scale-in — set
AWS_LIFECYCLE_DRAIN=1oninbound,sbc-sip-sidecarandfeature-server, and onsbc-rtpengine-sidecaron dedicated RTP servers. Allow the instance rolesautoscaling:DescribeAutoScalingInstancesandautoscaling:DescribeLifecycleHooks(in addition toautoscaling:CompleteLifecycleAction).AWS_SNS_TOPIC_ARNno longer turns on draining, and the SNS topics and their open ports can be removed. - API server behind a reverse proxy — set
JAMBONES_TRUST_PROXY=1and have the proxy sendX-Forwarded-For. Without both, all callers share one rate-limit counter. - Recordings — the API server’s built-in recording websocket is removed;
recordings are handled by
upload_recordings. - Audio capture opt-out — the Privacy settings appear in hosted portal
builds with
JAMBONES_OPT_OUT_VOIPMONITOR=true. The opt-out only takes effect if voipmonitor is configured withnorecord-header = yes.
Component Versions
Availability
jambonz.cloud - available now
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 for mini, medium and large, with community gallery images for all nine server types in 30 regions (amd64) and 29 regions (arm64).
self-hosted on GCP - available now: Terraform modules for mini, medium and large, with public images for all nine server types (GCP images are global), amd64 and arm64.
self-hosted Debian packages - available now: install jambonz-mini 11.1.6 from the
jambonz apt repository (Debian 12 bookworm), amd64 and arm64.