← Back to Blog

VoiceBlender v1.0.0: Not a Pet Project Any More

releasesiphigh-availabilitykubernetessrtpdraindocs

VoiceBlender v1.0.0 is out. It is a big step for the project.

Where VoiceBlender is today

VoiceBlender started as my own pet project. That is no longer true.

  • Companies use it. Large and small companies now run VoiceBlender for real calls.
  • People keep cloning it. The GitHub repository now gets up to 4,000 unique clones a day.
  • The Docker image is popular. It has been downloaded more than 8,000 times.
  • Other people write the code too. Several people now send changes. In this release, more than half of the commits came from people other than me.

Thank you to everyone who uses it, reports bugs and sends pull requests.

Built for more than one server

When you run VoiceBlender on several servers behind a SIP proxy or a load balancer, you need to update or restart one server without dropping calls. Most of this release is about that.

Drain mode

Drain mode takes one server out of service without cutting any calls.

  • New calls are refused with 503. A SIP proxy then sends them to another server straight away.
  • Calls that are already up keep going until they end on their own.
  • /health/ready starts to fail, so a load balancer stops sending new work. /health/live stays OK, so the server is not killed.
  • SIP trunks that register with a provider unregister, so the provider stops sending calls here. If you cancel the drain, they register again.

Start a drain from the API:

curl -X POST http://localhost:8080/v1/drain -d '{"timeout_seconds": 1800}'

Or send the process SIGUSR1. Then it drains and exits once the last call is over. DELETE /v1/drain cancels a drain.

You can set a time limit. If calls are still up when it runs out, the server shuts down cleanly anyway.

A clean shutdown

SIGTERM, which Docker and Kubernetes send when they stop a container, now runs a proper shutdown in a fixed order:

  1. Every call ends normally. The other side gets a BYE.
  2. Recordings are finished and uploaded.
  3. SIP trunks unregister.
  4. Waiting webhooks and VSI events are delivered.
  5. Then the process exits.

Before, a stop could lose the end of a recording or the last events. Now nothing is left half done. SHUTDOWN_TIMEOUT_SECONDS limits how long this takes. The configuration guide has a Kubernetes example.

Know what each server is doing

A new event, instance.status_changed, fires when a server changes state: starting, ready, draining, drained or stopping. It tells you how many calls and rooms are still on that server. Your application can use it to stop sending work to a server that is going away.

Calls that hold up better

  • Long calls through proxies. SIP session timers keep a call alive by refreshing it now and then. They now work correctly in both directions, so long calls through proxies and SBCs are no longer dropped by mistake.
  • Lost packets. On PCMU, PCMA and Opus calls, a lost audio packet is now filled in with a smooth guess instead of a click or a gap.
  • Network errors. A failed send of one audio packet no longer stops the audio on that call.
  • Recordings. A room recording keeps going when one person hangs up, and multi-channel room recordings no longer lose a person who hangs up.
  • Stability. We fixed several race conditions between a re-INVITE and the audio code.

Encrypted audio

SIP calls can now encrypt their audio with SDES-SRTP, including SRTCP. You can make it optional or required for the whole server, or turn it on for a single call. GET /v1/legs/{id}/streams shows whether each stream is encrypted.

Also new since v0.15.0

  • Opus bitrate. VoiceBlender now sets the Opus bitrate and follows the limits the other side asks for.
  • Clearer WebRTC errors. A failed WebRTC handshake now always ends with the reason dtls_failed.
  • Deepgram and ElevenLabs. You can choose a regional endpoint and opt out of data retention.
  • AMR. Better silence handling (DTX and comfort noise).
  • Security fixes and library updates.

New documentation

The documentation has been rebuilt from scratch.

  • Always up to date. The docs are made straight from the code, so they match what the server really does.
  • Everything in one place. There are three tabs: the REST API, the real-time WebSocket API (VSI) with all its commands and events, and every configuration setting with its default value.
  • Try it in the browser. You can send test requests from the page and copy ready-made code examples in many languages.
  • Easy to find things. Endpoints are grouped by topic, for example “Inbound calls”, “Transfer” or “Recording & SIPREC”.
  • One page per release. A version menu shows the docs for each release, so you can read the docs for the version you run.

Old links to the previous docs pages still work. They now open the new reference.

Upgrading

go get github.com/VoiceBlender/voiceblender@v1.0.0
go build -o voiceblender ./cmd/voiceblender

The Python SDK is on PyPI too:

pip install voiceblender

Full changelog: v0.15.0…v1.0.0.

Check the API Documentation for drain mode, health checks and SRTP.

Questions or feedback? Come say hi in our Discord - issues and PRs very welcome.