Brisvox changelog

The latest release is v0.1.3, published on . The Rust crates, the Python package and the npm packages are released together under one version number.

This page follows the GitHub releases of the project and updates itself. The format follows Keep a Changelog, and the project follows Semantic Versioning.

Last updated:

v0.1.3Latest

Release on GitHub

Added

  • Noise suppression of the caller's audio, opt in per agent: [audio] noise_suppression = "rnnoise" (default "off") filters the audio before the voice activity detector and the turn detector hear it, with RNNoise through the pure-Rust nnnoiseless crate (BSD-3-Clause; cargo feature denoise-rnnoise, not in the default features of brisvox-server; a server built without it refuses the file at the key's position). The filter runs on the inference pool inside the VAD batch job, costs about 0.35 % of one core per call (0.45 % on an 8 kHz call) and adds 20 to 25 ms of delay to what the VAD hears; it never delays or drops the caller's audio, and a failing or overloaded filter lets the audio through unfiltered. denoise_stt = true also sends the filtered audio to the speech-to-text provider (default false: independent studies found that denoised audio often raises the word error rate). Library: brisvox_inference::denoise (the streaming Denoiser trait, VadStream::with_denoiser), SessionBuilder::with_denoise, DenoiseStatsSnapshot. A call below a filter's minimum sample rate runs without it (logged once), and a filter that cannot be built leaves the call with the plain audio and a warning; both are counted in brisvox_engine_events_total{event="denoise_bypassed_narrowband" | "denoise_build_failures" | "denoise_errors" | "denoise_feed_dropped"}. Measured with 34 real noise recordings (bench/results/denoise-*.md): it does no harm on clean audio but did not reduce false speech starts on wideband audio and shifted Smart Turn verdicts on telephone audio, so it stays off by default and is meant to be tested on one's own calls. brisvox-bench denoise cpu|quality|transcribe measures the CPU per call, the false barge-ins and turn verdicts under licensed noise (fixtures/noise, public domain), and streams the same mixes to a sherpa-onnx server for the word error rate.

  • Reload the agent file without a restart. Edit the agent file and POST /admin/reload on the health listener (or kill -HUP the server on Unix, or set BRISVOX_AGENT_RELOAD_POLL_MS to have the file polled for changes): new calls run the new definition, calls in progress keep theirs until they end. The new file is validated like the start validates it, MCP servers included; an invalid file, an MCP server that cannot be reached or a change that needs a restart (the [recording] table, a provider limit or Smart Turn model the start did not set up) leaves the old definition running and is reported with the file position. The MCP servers and provider connections of the old definition close when its last call ends. The endpoint answers JSON and needs BRISVOX_ADMIN_TOKEN (a separate secret from the metrics token) unless the health listener is on loopback; off a loopback address without a token it is off. The file poll asks for a reload once per change, so a refused file is not retried every interval. The server now handles SIGHUP (a closed terminal reloads instead of ending it). New metrics brisvox_agent_reloads_total{result} and brisvox_agent_generation, one log line per reload. See docs/OPERATIONS.md ("Reloading the agent") and docs/CONFIG.md. For embedders, the new brisvox_server::reload module has the pieces (SwapRunner, RunnerRecipe, AgentSet, ReloadHub).

  • Language switching mid-call. An agent file can list several languages with [languages] and [[languages.speak]] (a voice, and optionally a TTS model, for each). The speech-to-text then listens to all of them for the whole call (Deepgram language=multi, AssemblyAI language detection, Gladia code switching, OpenAI languages), and when the caller speaks another listed language on two utterances in a row (configurable, or once when the provider is confident) the next reply is in that language, with that language's voice, without a new call. The language comes from the STT provider's report (new language field on SttEvent::Transcript) or, when it reports none, from a text detector. The reply language is pinned per turn with a Reply in <language>. line, never in the middle of a response. A file that lists a language its TTS provider cannot speak (Deepgram Aura, Kokoro and Turkish) is refused at its position. New metric brisvox_language_switches_total, new agent protocol event language.switched, new TtsConfig::language and TtsConfig::model for adapters, example examples/agents/multilingual-tr-en.toml. See docs/CONFIG.md, "Languages".

Changed

  • Cheaper turn detection is the default: the Smart Turn model runs with each GELU fused into one operation (BRISVOX_SMART_TURN_GRAPH=fused; about 40% less memory traffic per analysis and more calls per core under load). In 0.1.2 it was opt-in because one early verdict on a test cut ended a turn the published graph did not; measured on the official test set, both graphs do that at the same small rate and end the same number of turns too soon (4000 clips with VNNI, 1000 without; bench/results/smart-turn-early-verdict-2026-10-31.md). BRISVOX_SMART_TURN_GRAPH=original runs the file as published.

  • Adapter API (library users): SttEvent::Transcript has a new field language: Option<SpokenLanguage> (the language the STT provider reported for a final segment; None when it reports none), and TtsConfig has two new fields, language and model (the language and TTS model a session should speak, None to keep the adapter's own). Code that builds these values with struct literals must add the fields; code that matches SttEvent::Transcript without .. must name the new field. The agent protocol gains the language.switched event (new LanguageSwitched message in the Python and TypeScript SDKs).

  • Background speech and echo handling, measured against the engine's real barge-in guards (bench/results/noise-handling-calls.md, docs/LATENCY.md, "Noise handling beyond a denoiser"):

    • [audio] noise_suppression = "hush" (experimental): Hush (Apache-2.0, pure Rust, trained for voice agents with competing speakers; cargo feature denoise-hush, not in the default features) as the filter in front of the VAD and the turn detector: 1 % of a core per call, 10 ms of delay; takes a second voice 15 to 20 dB under the caller and the bot's echo out of the barge-ins in many calls, no extra delay for a real barge-in. Its gain depends on the input level (it removes the speech of a quiet caller and distorts 8 kHz phone audio), so a call below 16 kHz runs without it (logged, counted) and the agent file refuses it with call_sample_rate below 16000. brisvox-bench denoise safety checks that a filter leaves clean speech at three caller levels alone. "rnnoise" measured worse than nothing for the VAD and is not in the default features either.
    • [interruption] echo_text_match: holds a barge-in whose words are the bot's own, just said (the bot's voice coming back); needs min_words >= 1. Removed the echo case completely in the measurements. A caller who repeats the bot's words back while it speaks is held like the echo until other words appear (the texts are the same); short commands (at most two words with "stop", "wait", "no", ...) always get through.
    • [vad] level_gate_db: speech may only start within that many dB of the caller's own level, so a quieter second voice cannot start a barge-in.
    • Library: brisvox_inference::Gated (level gates and a noise-floor estimate over any speech model), brisvox_inference::denoise::blend::Blended (dry/wet mix for any denoiser), brisvox_agent::{TurnMachine::set_echo, TurnMachine::set_interruption, echoes_bot, is_bot_echo, NoiseAdapt}. brisvox-bench denoise calls drives the real turn machine and policy over calls with background speech (LibriSpeech, fixtures/interferers), bot echo and noise.

Fixed

  • tokio 1.53.3 in the lock file (RUSTSEC-2026-0386, AsyncFd::new and related constructors violate I/O safety). brisvox does not call those constructors; the crates take any tokio 1.x, so their users pick up the fix with cargo update -p tokio.

v0.1.2

Release on GitHub

Added

  • Keyless local stack: docker compose -f deploy/local/compose.yaml up starts a voice agent you talk to in the browser with no API key and no cloud: brisvox-server, a sherpa-onnx streaming recognizer, Ollama with qwen2.5:1.5b and a Kokoro-FastAPI voice, all on the CPU, with health checks, a first-start model pull, a Llama 3.2 3B variant for faster machines and a build-from-source override. A weekly workflow plays one recorded conversation at the stack. See deploy/local/README.md and docs/DEPLOY.md.
  • MCP tools for agents: an agent file can name MCP (Model Context Protocol) servers, as a program the server starts (stdio) or a streamable HTTP endpoint, with [[mcp.servers]]. The tools on a server's allowlist are offered to the LLM like the server's own tools, with the same guarantees: tools marked read_only in the file may run speculatively and are dropped on a barge-in, every other tool never runs speculatively and never twice for one call, confirm holds a tool for the caller's confirmation, and the engine's idempotency key travels in the request's _meta. A stdio server that exits is restarted with backoff, and calls fail at once meanwhile; the child process gets only its own environment, never the provider keys. Plain http:// is accepted only to a private-network host unless the entry sets allow_insecure, and max_in_flight (default 32) bounds the calls in flight per server. Cargo feature mcp (in the prebuilt binaries and the image). See docs/TOOLS.md ("MCP tools"), docs/CONFIG.md and the example examples/agents/mcp/docs-en.toml.
  • Grafana dashboard for the server's Prometheus metrics: deploy/grafana/brisvox.json, one row per operator question (healthy now, fast, do barge-ins work, why slow or refusing, providers failing, cost). Tested with Grafana 12.1 (UI import and file provisioning). A test checks every metric, label and label value it queries against the real /metrics output, and rejects quantiles without le and counters without a rate.
  • Cheaper turn detection, opt-in: BRISVOX_SMART_TURN_GRAPH=fused runs the Smart Turn model with each GELU fused into one operation (rewritten in memory from the pinned file): about 40% less memory traffic per analysis and more calls per core under load; on 4000 clips of the official Smart Turn v3.2 test set the verdicts matched the published graph on 3997 and the accuracy did not drop. It is not the default because on one abrupt cut of the test fixtures its early verdict ended a turn that the published graph's did not (docs/providers/smart-turn.md, "Fused GELU").

Changed

  • Benchmarks: the capacity numbers in the README and the book now come from cloud containers (AMD EPYC 9655P, 4 vCPU / 4 GB and 8 vCPU / 8 GB) with every call going through each system's own provider client code to protocol-faithful fake providers over TLS, and the real Silero VAD and Smart Turn models: brisvox 250 and 470 calls within the latency SLO against 1 and 10 for Pipecat 1.12.0, and 430 and 780 without a failed call or lost turn against 10 and 10. The local harness runs the same comparison: PROVIDERS=fakes MODELS=1 bench/capacity_linux.sh.

Fixed

  • TLS certificate checks: rustls-webpki 0.103.16 and webpki-roots / webpki-root-certs 1.0.11, so provider connections no longer trust the distrusted emSign intermediates and reject distrusted intermediates in a chain (RUSTSEC-2026-0381, RUSTSEC-2026-0382, RUSTSEC-2026-0383).
  • sherpa-onnx: a hang-up ends the stream with Done and lets the server close the socket, so the Python server no longer logs a failed connection after every call.

v0.1.1

Release on GitHub

Added

  • Usage and cost per call: the engine counts the provider billing units of every call (LLM tokens including discarded speculative responses, STT audio seconds sent, TTS characters sent) and the server delivers one per-call summary (duration, turns, average and worst stage latency, units with the configured provider and model) as the agent link event session.usage before session.end, as the summary of the transcript's end line, as one call summary log line and as the Prometheus families brisvox_llm_tokens_total, brisvox_stt_audio_seconds_total, brisvox_tts_characters_total and brisvox_call_duration_seconds; bot.llm_stopped carries the tokens of each turn, and the Python and TypeScript agent SDKs have a typed session.usage event. Units, not money (docs/OPERATIONS.md, "Usage and cost").
  • Learned admission ceiling (BRISVOX_OVERLOAD_CEILING, on by default). Pool pressure shows only when callers speak, so a fast ramp used to be admitted far past capacity (on 4 vCPU, about 730 calls held for a capacity of about 250, every call at 3.7 s p50). The server now learns how many calls it serves well from the pressure signals and refuses new calls (503) above that, without touching running calls. In a load test with 900 callers dialing into 4 vCPU: 120 to 400 calls held at 0.9 to 1.5 s p95, none failed.
  • Cautious start: a fresh server admits 16 calls and doubles per calm window until the pool is first under pressure, so a rush right after a restart cannot overshoot.
  • A second agent per server (BRISVOX_WS_AGENT_FILE): the generic websocket listener runs its own agent, sharing the process's pools and admission control.
  • BRISVOX_DEFAULT_AGENT_FILE: the container image's default agent path, used only when neither BRISVOX_AGENT_FILE nor BRISVOX_AGENT_DEFINITION is set.

Changed

  • Library API (brisvox-engine, brisvox-agent-link, brisvox-server): EngineEvent::BotLlmStopped has a new usage field (patterns need ..); CallGuard::new takes a summarizer argument; Metrics::observe_call and CallObserver::call_ended take a summary argument; observability::metrics::Histogram::new is no longer const (use Histogram::with_bounds for other buckets); new AgentLink::close_with (close is unchanged).
  • BRISVOX_MAX_SESSIONS defaults to 256 per core (at least 256) instead of 256.
  • Pool pressure also watches the Smart Turn executor (its queue waits): a late end-of-turn verdict delays the reply as much as a late VAD window. /readyz reports overloaded at the learned ceiling and when the Smart Turn executor is behind, so a load balancer sends calls elsewhere.

Fixed

  • An admitted call could stay silent on a saturated pool: the output resampler is now built off the pool when the pool is full or late.
  • The container image set BRISVOX_AGENT_FILE, so BRISVOX_AGENT_DEFINITION alone (the way to pass an agent on a platform that cannot mount a file) failed at startup.
  • The release workflow's npm publish step (./pkg/*.tgz): 0.1.0 reached GitHub, GHCR and PyPI only; crates.io and npm start at 0.1.1.

v0.1.0

Release on GitHub

First public release.

Added

  • Engine. A Rust-native real-time voice pipeline, one isolated session per call: transport pump, VAD, STT, turn driver, LLM runner, TTS and transport out. Turn epochs drop stale frames everywhere, so an interruption never leaves old audio or text behind. After an interruption the LLM context holds only what was actually played (played-text truncation). Every channel is bounded, every network await has a timeout, every task is cancellable, and one session's panic or provider failure cannot touch another.
  • Server. brisvox-server, a multi-tenant session server: one TOML or YAML file per agent (brisvox-server check <file> validates it offline with line and column for every problem), health and readiness endpoints, Prometheus metrics, OpenTelemetry traces (feature otlp), admission control, graceful drain, recordings encrypted at rest with age (feature recording-encryption), a container image and a compose file.
  • Transports. Generic websocket, Twilio Media Streams, Telnyx media streaming, RTVI (Pipecat's client SDKs), WebRTC (feature webrtc, WHIP-style signaling, Opus) and a SIP trunk listener (feature sip). G.711 mu-law and A-law, any call rate with any provider rate.
  • Providers. STT: Deepgram, AssemblyAI, Gladia, sherpa-onnx (on your machine) and servers that speak the OpenAI Realtime transcription protocol. LLM: OpenAI, Anthropic, Gemini and any OpenAI-compatible endpoint (Groq, Together, Fireworks, vLLM, Ollama, OpenRouter). TTS: Cartesia, ElevenLabs, Deepgram Aura and any server of the OpenAI /v1/audio/speech shape. Speech-to-speech: OpenAI Realtime and Gemini Live. Per-stage failover, circuit breakers and retry budgets.
  • On-device models. Silero VAD (compiled into the binary, MIT) and Smart Turn v3.2 end-of-turn detection on ONNX Runtime (feature smart-turn-ort), on a dedicated inference pool.
  • Conversation. Adaptive turn taking, interruption policy with backchannel lists, speculative LLM start, tool calls with read_only and side_effect kinds (a side-effect tool never runs speculatively and never runs twice), filler phrases for slow tools, DTMF in and out, call transfer, answering machine handling. English and Turkish.
  • SDKs. The Python package brisvox-agent and the npm package @brisvox/agent for an out-of-process agent (your tools and application logic in your own process, the server runs the call), and @brisvox/web, a dependency-free browser module for a "Talk to us" button: WebRTC first, automatic WebSocket fallback, microphone, playback, barge-in and clean hang-up.
  • Rust crates. brisvox-core, brisvox-audio, brisvox-inference, brisvox-agent, brisvox-providers, brisvox-sim, brisvox-transport, brisvox-engine, brisvox-agent-link and brisvox-server on crates.io; brisvox_server::app::run_with_tools is a whole main for a server with your own Rust tools.
  • Umbrella packages. The plain name brisvox is a real package on all three registries, so cargo add brisvox, pip install brisvox and npm i brisvox work next to the specific names. The crate re-exports brisvox-server, the tools API (brisvox-agent), the provider adapters and the core types and forwards the server's Cargo features (brisvox::run_with_tools(...) is a whole main). The PyPI package depends on brisvox-agent of the same release and re-exports its API (from brisvox import Agent). The npm package depends on @brisvox/agent and @brisvox/web and re-exports them (brisvox is the agent SDK for Node, brisvox/web the browser client).
  • Benchmarks. A harness for voice-to-voice latency and capacity with real and simulated providers, with the raw results and the method committed under bench/.
  • Documentation. The book (architecture, quickstart, configuration, telephony, operations, benchmarks, protocols and provider notes) and llms.txt / llms-full.txt for AI coding agents.
  • Release. Prebuilt brisvox-server binaries for Linux (x86_64, aarch64), macOS (aarch64) and Windows (x86_64) with checksums, and a multi-architecture image at ghcr.io/bilal07karadeniz/brisvox-server.