HomeLocal speech API

Developer speech infrastructure

Add local speech without embedding speech models.

Call one stable local bridge for synthesis, transcription, streaming, cancellation, and capability discovery. Use one-shot commands for simple automation or the warm Unix-socket daemon for repeated requests.

View OSS source Read integration source

MIT-licensed coreStructured stdoutPrivate Unix socketNo network service required

Bridge contractOSS + Pro

Core commands

avrvx --synthesize --text-file reply.txt --output reply.wav
avrvx --transcribe recording.ogg
avrvx --capabilities
avrvx --daemon

The bridge is included in both editions. Pro capabilities are discovered at runtime.

CLI / socketTransport

JSONDiscovery

PCM16Streaming

Four jobs, one executable

Use the interface that matches your application lifecycle.

The bridge keeps diagnostics on stderr and machine-readable results on stdout, making it safe to compose in shells and host processes.

Synthesize

Write a mono PCM WAV file, stream headerless PCM16 to stdout, or select an installed voice and one-call speed override.

Transcribe

Pass an existing audio recording through local Faster Whisper and receive plain transcript text with meaningful exit status.

Discover capabilities

Read edition, license state, version, TTS engines, voices, STT model, daemon state, and feature flags as JSON.

Serve warm requests

Load speech resources once and accept newline-delimited JSON requests over an owner-only Unix socket.

File-oriented by default

Simple commands for scripts, cron jobs, and occasional work.

Use stdin or a protected text file for sensitive material instead of a command-line argument visible in the process list.

WAV outputs are created mode 0600
Output directories must already exist
Transcription exits non-zero when no speech is found
SIGTERM produces a distinct cancellation status
Diagnostics stay on stderr

One-shot examples

# Prefer stdin for sensitive text
printf '%s' "Deployment complete" |       avrvx --synthesize --text - --output /tmp/reply.wav

# Transcribe anything your ffmpeg path can decode
avrvx --transcribe /tmp/voice-message.ogg

# Parse capability JSON
avrvx --capabilities | jq '.tts.engines, .features'

Use the sample rate reported by the bridge rather than assuming every voice uses the same rate.

Low-latency repeated use

Keep models warm behind a private local socket.

The daemon is an optimization, not a second deployment mode. Clients can try it first and transparently fall back to normal processes.

Socket: $XDG_RUNTIME_DIR/avervox/bridge.sock
Permissions: Mode 0600 inside a mode-0700 directory
Protocol: Newline-delimited JSON; optional framed streaming for synthesis
Operations: capabilities, synthesize, transcribe, cancel, ping
Failure mode: Client fallback to one-shot avrvx invocation

Streaming PCM

avrvx --synthesize       --text-file reply.txt       --output - |       aplay -f S16_LE -r 22050 -c 1

This example uses a common Piper medium-voice rate. Production clients should read the announced format instead of hard-coding 22050 Hz.

Host integrations

Ship one provider package for OSS and Pro.

Capability discovery allows the host to expose the features actually installed rather than requiring separate builds.

Hermes Agent

Python provider with TTS, STT, streaming, cancellation, and warm-daemon support.

Hermes guide

OpenClaw

Native plugin published through npm and ClawHub for streaming local TTS.

OpenClaw guide

Integration template

Reuse the same pattern for another host: detect capabilities, prefer daemon, fall back to process, preserve cancellation.

Open repository

Bridge questions

Design for the failure modes, not only the happy path.

Is the daemon reachable over the network?

No. It uses a local Unix-domain socket under the current user’s runtime directory. Pro’s separate LAN voice server is an explicit feature with different configuration.

Can stdout always be parsed?

Bridge commands keep diagnostics on stderr. Capability discovery uses JSON on stdout; transcription uses plain text; synthesis outputs a path or the requested PCM stream.

How should sensitive text be passed?

Prefer --text-file or --text - with stdin. Command-line arguments can be visible to other local users through the process list on some systems.

How do I support both OSS and Pro?

Call --capabilities at startup or when configuration changes. Use the reported engines, voices, edition, license state, and feature map to build the host UI.

Can the host cancel synthesis?

Yes. One-shot processes react to termination, and the daemon protocol includes a cancel operation for active requests.

Start with the workflow

Build on the local speech layer instead of rebuilding it.

Use the MIT-licensed bridge in scripts and applications, then let users unlock Pro voices and automation without changing your integration.

View OSS source Get Pro — $39