macos-speech-server

Local, private speech-to-text (STT) and text-to-speech (TTS) server for macOS with OpenAI-compatible and Home Assistant (Wyoming) support.

Runs entirely on-device using Apple's Neural Engine via FluidAudio -- no cloud services, no API keys, no data leaves your machine. Models are loaded once at startup and served to any device on your network, so a single Mac with Apple Silicon can handle transcription and speech for your entire household.

Two interfaces, one server:

OpenAI-compatible HTTP API -- drop-in replacement for OpenAI audio endpoints (/v1/audio/transcriptions, /v1/audio/speech)
Wyoming protocol (TCP, default port 10300) -- native Home Assistant voice pipeline integration

Requirements

macOS 14+
Swift 6.2+
Apple Silicon recommended (Neural Engine acceleration)

Quick start

swift build
swift run speech-server

On first launch, ASR and TTS models are downloaded automatically. This takes several minutes but only happens once; subsequent starts are fast.

The server listens on http://localhost:8080 by default. The Wyoming protocol server listens on TCP port 10300 by default.

Configuration

All server settings can be customised via a YAML config file. Create speech-server.yaml in the working directory (a fully-commented example is included in the repo):

log_level: notice     # trace | debug | info | notice | warning | error | critical

servers:
  http:
    host: 127.0.0.1       # use your LAN or Tailscale IP to accept connections from other devices
    port: 8080
    upload_limit_mb: 500
  wyoming:
    host: 127.0.0.1       # can differ from http.host; set independently
    port: 10300           # TCP port for Wyoming protocol (Home Assistant). 0 = disabled.

stt:
  engine: parakeet      # Currently only: parakeet (NVIDIA Parakeet TDT via FluidAudio)
  parakeet:
    model_version: v3   # v3 = multilingual (25 langs, default), v2 = English-only

tts:
  engine: pocket_tts    # pocket_tts (default) | avspeech | kokoro

  # AVSpeech settings (only used when engine: avspeech)
  # avspeech:
  #   default_voice: Samantha   # Short name or full identifier; nil = system locale default
  #   sample_rate: 22050        # Native AVSpeech output rate (Hz)

  # Kokoro TTS settings (only used when engine: kokoro)
  # kokoro:
  #   default_voice: af_heart   # Any Kokoro voice ID (e.g. af_heart, am_adam); default af_heart

All fields are optional — omitted fields use the defaults shown above.

TTS engines

Three TTS engines are available:

Engine	`engine:` value	Voices	Sample rate	Downloads	Notes
FluidAudio PocketTTS	`pocket_tts`	`alba` only	24 kHz	~200 MB on first start	Default
macOS AVSpeech	`avspeech`	150+ system voices	22050 Hz	None (ships with macOS)	Instant startup
FluidAudio Kokoro	`kokoro`	50 voices, 8 languages	24 kHz	~300 MB on first start	High quality

`pocket_tts` (default)

Uses FluidAudio's PocketTTS model. Only the alba voice is available. Models are downloaded on first start and cached at ~/Library/Application Support/FluidAudio.

`avspeech` — macOS built-in voices

Uses macOS's AVSpeechSynthesizer — no model downloads, instant startup, 150+ voices across dozens of languages. Audio is synthesised at 22050 Hz mono (16-bit PCM).

tts:
  engine: avspeech
  avspeech:
    default_voice: Samantha   # Optional — nil uses the system locale default

List all available voices with:

say --voice '?'

The short name (e.g. Samantha, Daniel, Karen) is used in API requests. Voice names are case-insensitive; full identifiers (e.g. com.apple.voice.enhanced.en-US.Samantha) also work.

Note: Siri voices are not accessible via public AVFoundation APIs and will not appear in the voice list. Personal Voice support (macOS 14+) is planned — see issue #13.

`kokoro` — FluidAudio Kokoro

Uses FluidAudio's Kokoro CoreML model. 50 voices across 8 languages (American English, British English, Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese, Mandarin Chinese), synthesised at 24 kHz. Models are downloaded on first start and cached at ~/.cache/fluidaudio/Models/kokoro.

tts:
  engine: kokoro
  kokoro:
    default_voice: af_heart   # Optional — default is af_heart (American English female)

American English voices (production-quality): af_alloy, af_aoede, af_bella, af_heart, af_jessica, af_kore, af_nicole, af_nova, af_river, af_sarah, af_sky, am_adam, am_echo, am_eric, am_fenrir, am_liam, am_michael, am_onyx, am_puck, am_santa.

Other language voices are experimental (not QA'd). Full voice list: use /v1/audio/speech with an invalid voice to see the available options listed in the error message, or query GET /v1/models via your client library.

Config discovery order

SPEECH_SERVER_CONFIG environment variable (path to a YAML file)
./speech-server.yaml in the current working directory
Built-in defaults (no file needed)

# Use an explicit config file via env var
SPEECH_SERVER_CONFIG=/etc/speech-server.yaml swift run speech-server

Environment variable overrides

Individual settings can also be overridden with environment variables:

Variable	Overrides	Example
`HTTP_HOST`	`servers.http.host`	`HTTP_HOST=192.168.1.50`
`HTTP_PORT`	`servers.http.port`	`HTTP_PORT=9090`
`WYOMING_HOST`	`servers.wyoming.host`	`WYOMING_HOST=192.168.1.50`
`WYOMING_PORT`	`servers.wyoming.port`	`WYOMING_PORT=0` (disables Wyoming)

Vapor's --hostname and --port CLI flags also work and take highest priority for the HTTP server.

Deployment

This project supports two launchd deployment models:

Model	Scope	Privileges	Persistence	Install script
LaunchDaemon	System-wide	`sudo` required	Starts at boot, survives logout	`deploy/install-daemon.sh`
LaunchAgent	Per-user	No `sudo`	Starts at user login	`deploy/install-agent.sh`

Option A: LaunchDaemon (persistent system service)

Recommended when you want the server always available on the machine.

Build and install:

sudo deploy/install-daemon.sh

Verify status:

sudo launchctl print system/com.local.speech-server

Watch logs:

sudo tail -f /var/log/speech-server/output.log /var/log/speech-server/error.log

What the installer does:

Creates dedicated service account _speech-server.
Installs binary to /usr/local/bin/speech-server.
Installs config to /etc/speech-server/speech-server.yaml (without overwriting existing config).
Installs LaunchDaemon plist to /Library/LaunchDaemons/com.local.speech-server.plist.
Optionally pre-populates FluidAudio cache from invoking user into the service account home.

Useful daemon commands:

# Restart daemon after config changes
sudo launchctl kickstart -k system/com.local.speech-server

# Stop daemon
sudo launchctl bootout system/com.local.speech-server

# Start daemon again
sudo launchctl bootstrap system /Library/LaunchDaemons/com.local.speech-server.plist
sudo launchctl enable system/com.local.speech-server
sudo launchctl kickstart -k system/com.local.speech-server

Uninstall daemon:

# Keep config/logs/service user
sudo deploy/uninstall-daemon.sh

# Remove everything including config/logs/model cache/service user
sudo deploy/uninstall-daemon.sh --purge

Option B: LaunchAgent (per-user service)

Recommended when you want user-session startup with no system-wide changes.

Build and install:

deploy/install-agent.sh

Verify status:

launchctl print gui/$(id -u)/com.local.speech-server

Watch logs:

tail -f "$HOME/Library/Logs/speech-server/output.log" "$HOME/Library/Logs/speech-server/error.log"

Useful agent commands:

# Restart agent after config changes
launchctl kickstart -k gui/$(id -u)/com.local.speech-server

# Stop agent
launchctl bootout gui/$(id -u)/com.local.speech-server

# Start agent again
launchctl bootstrap gui/$(id -u) "$HOME/Library/LaunchAgents/com.local.speech-server.plist"
launchctl enable gui/$(id -u)/com.local.speech-server
launchctl kickstart -k gui/$(id -u)/com.local.speech-server

Uninstall agent:

# Keep config/logs
deploy/uninstall-agent.sh

# Remove user config/logs too
deploy/uninstall-agent.sh --purge

Deployment files

The deploy/ directory contains ready-to-use templates and scripts:

deploy/com.local.speech-server.plist -- LaunchDaemon template
deploy/com.local.speech-server.agent.plist -- LaunchAgent template (__HOME__ placeholders are replaced by installer)
deploy/install-daemon.sh and deploy/uninstall-daemon.sh
deploy/install-agent.sh and deploy/uninstall-agent.sh

API

All endpoints are available at both /audio/* and /v1/audio/* (OpenAI compatibility).

Speech-to-Text

POST /v1/audio/transcriptions
Content-Type: multipart/form-data

Field	Type	Required	Description
`file`	File	Yes	Audio file (max 500 MB)
`model`	String	No	Model name (e.g. `whisper-1`)
`language`	String	No	ISO-639-1 language code
`prompt`	String	No	Context hint for transcription
`response_format`	String	No	`json` (default), `text`, or `verbose_json`; `srt`/`vtt` return 400
`temperature`	Double	No	Sampling temperature, 0.0-1.0

Supported audio formats: WAV, MP3, M4A, FLAC, AIFF, OGG. Files without a recognised extension are identified automatically via magic bytes.

No API key is required. If your client sends an Authorization header it is silently ignored.

The verbose_json response includes a segments array and real duration from the ASR engine, matching the OpenAI API shape:

{
  "task": "transcribe",
  "language": "en",
  "duration": 1.54,
  "text": "Hello world.",
  "segments": [{ "id": 0, "seek": 0, "start": 0.0, "end": 1.54, "text": "Hello world.", ... }]
}

Example:

curl -X POST http://localhost:8080/v1/audio/transcriptions \
  -F file=@recording.wav -F model=whisper-1

curl -X POST http://localhost:8080/v1/audio/transcriptions \
  -F file=@recording.wav -F model=whisper-1 -F response_format=verbose_json

Text-to-Speech

POST /v1/audio/speech
Content-Type: application/json

Field	Type	Required	Description
`model`	String	Yes	Model name (e.g. `tts-1`)
`input`	String	Yes	Text to synthesize (max 4096 chars)
`voice`	String	No	Voice name (default: engine default). See TTS engines.
`response_format`	String	No	`wav` (default) or `pcm`
`speed`	Double	No	Playback speed, 0.25-4.0 (default: 1.0)

The response is streamed: audio begins arriving before synthesis is complete, sentence by sentence. WAV responses include a standard 44-byte header (with unknown-size placeholders) followed by 16-bit PCM; PCM responses are raw 16-bit bytes. The sample rate depends on the active TTS engine (24 kHz for pocket_tts and kokoro, 22050 Hz for avspeech).

Example:

curl -X POST http://localhost:8080/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model":"tts-1","input":"Hello, world!"}' \
  --output speech.wav

# AVSpeech engine with a specific voice
curl -X POST http://localhost:8080/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model":"tts-1","input":"Hello, world!","voice":"Samantha"}' \
  --output speech.wav

Compatible apps

The HTTP API is compatible with any app or library that supports a configurable OpenAI base URL. No real API key is needed -- the server ignores Authorization headers, so enter any non-empty string.

MacWhisper

MacWhisper has built-in support for custom transcription providers:

Open MacWhisper Preferences
Go to the Provider tab and choose Custom
Set the API URL to http://<host>:<port>/v1/audio/transcriptions (default localhost:8080)
Enter any string as the API Key (e.g. local)

Audio is sent directly to the endpoint; transcription happens entirely on-device with no round-trip to the cloud.

Other apps

Any tool that supports a configurable OpenAI base URL should work out of the box: set the base URL to http://<host>:<port> (default localhost:8080) and use any string as the API key. This includes the official OpenAI Python and JavaScript SDKs, and similar tools.

Accessing from other machines

By default the server binds to 127.0.0.1 and is only reachable locally. To serve requests from other devices -- another Mac, a phone, a Home Assistant instance -- bind to a reachable address and make sure the ports are accessible.

Tailscale (recommended)

Tailscale gives every device a stable private IP with no port-forwarding or firewall rules, and works across different networks (home, office, mobile). Both the HTTP API port and the Wyoming port are plain TCP; Tailscale handles encryption transparently.

Recipe:

Install Tailscale on the Mac running the server and on any device that needs access.
Note the Mac's Tailscale IP (e.g. 100.x.y.z) from the menu-bar icon.
Bind the server to that IP in speech-server.yaml:

servers:
  http:
    host: 100.x.y.z   # your Mac's Tailscale IP
  wyoming:
    host: 100.x.y.z   # your Mac's Tailscale IP
    port: 10300

Point your client at http://100.x.y.z:8080 (HTTP API) or 100.x.y.z:10300 (Wyoming).

Local network

Find your Mac's LAN IP in System Settings > Network, select your active connection (Wi-Fi or Ethernet), and note the IP address (e.g. 192.168.1.50). Bind the server to that address:

servers:
  http:
    host: 192.168.1.50   # your Mac's LAN IP
  wyoming:
    host: 192.168.1.50   # your Mac's LAN IP
    port: 10300

Use that same IP in your client configuration. Note that LAN IPs can change when devices reconnect; consider assigning a DHCP reservation in your router, or use Tailscale for a stable address.

Home Assistant

macos-speech-server speaks the Wyoming protocol, enabling fully on-device STT and TTS for Home Assistant voice pipelines via the Wyoming integration.

A single TCP port (default 10300) handles both STT and TTS -- Home Assistant discovers both capabilities automatically.

Network setup

Home Assistant typically runs on a separate machine, so the Wyoming port must be reachable from it. See Accessing from other machines above for Tailscale and LAN options -- in either case, set servers.http.host and servers.wyoming.host to your Mac's IP (or use the HTTP_HOST and WYOMING_HOST environment variables) so both ports are reachable from HA.

Adding the Wyoming integration in Home Assistant

The integration must be added manually (zeroconf/auto-discovery is not supported):

Go to Settings > Devices & Services
Click Add Integration
Search for Wyoming Protocol
Enter the host (IP address of the Mac running macos-speech-server) and port (default 10300)
Home Assistant discovers both STT and TTS capabilities on that single port

Using in a voice pipeline

Go to Settings > Voice Assistants
Create a new pipeline or edit an existing one
Select macos-speech-server for the Speech-to-text and/or Text-to-speech step

Streaming TTS (lower latency, audio starts playing before synthesis is complete) is supported in Home Assistant 2025.07 and later.

Project structure

speech-server.yaml.example         # Example config (all defaults); copy to speech-server.yaml to customise
                                   # speech-server.yaml is gitignored (may contain private IPs)
Sources/speech-server/
  Entrypoint.swift                 # Application entry point
  configure.swift                  # Middleware and service setup
  routes.swift                     # Route registration
  ServerConfig.swift               # YAML config loading + Vapor DI
  Controllers/
    TranscriptionController.swift  # STT endpoint
    SpeechController.swift         # TTS endpoint
  Services/
    STTService.swift               # STT protocol + DI
    FluidSTTService.swift          # FluidAudio ASR implementation (parakeet engine)
    AudioFormatDetection.swift     # Magic-byte audio format detection
    TTSService.swift               # TTS protocol + DI
    FluidTTSService.swift          # FluidAudio PocketTTS implementation (pocket_tts engine)
    AVSpeechTTSService.swift       # macOS AVSpeechSynthesizer implementation (avspeech engine)
    KokoroTTSService.swift         # FluidAudio Kokoro implementation (kokoro engine)
    PCMConversion.swift            # Shared Float32→Int16 PCM conversion and WAV builder
    SentenceDetection.swift        # Shared sentence splitting for TTS
  Middleware/
    RequestLoggingMiddleware.swift  # Logs method, path, status code
    OpenAIErrorMiddleware.swift    # OpenAI-format error responses
  Models/
    TranscriptionResponse.swift
    SpeechRequest.swift
    OpenAIError.swift
  Wyoming/
    WyomingEvent.swift             # Protocol event model
    WyomingFrameDecoder.swift      # Wire format parser
    WyomingNIOHandler.swift        # NIO channel handler
    WyomingServer.swift            # TCP server bootstrap
    WyomingSession.swift           # Session state machine (STT + TTS)
    WyomingWAVWriter.swift         # PCM-to-WAV for STT handoff

Contributing

Contributions are welcome. All changes go through a pull request — see CONTRIBUTING.md for the development workflow, code style, and PR guidelines.

Swift code is formatted with swift format (ships with Swift 6.2). A pre-commit hook is provided; install it with scripts/install-hooks.sh.

License

AGPL-3.0 -- see LICENSE.

Name		Name	Last commit message	Last commit date
Latest commit History 63 Commits
.github/workflows		.github/workflows
Sources/speech-server		Sources/speech-server
Tests		Tests
deploy		deploy
scripts		scripts
.gitignore		.gitignore
.swift-format		.swift-format
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
Package.resolved		Package.resolved
Package.swift		Package.swift
README.md		README.md
speech-server.yaml.example		speech-server.yaml.example

Folders and files

Latest commit

History

Repository files navigation

macos-speech-server

Requirements

Quick start

Configuration

TTS engines

pocket_tts (default)

avspeech — macOS built-in voices

kokoro — FluidAudio Kokoro

Config discovery order

Environment variable overrides

Deployment

Option A: LaunchDaemon (persistent system service)

Option B: LaunchAgent (per-user service)

Deployment files

API

Speech-to-Text

Text-to-Speech

Compatible apps

MacWhisper

Other apps

Accessing from other machines

Tailscale (recommended)

Local network

Home Assistant

Network setup

Adding the Wyoming integration in Home Assistant

Using in a voice pipeline

Project structure

Contributing

License

About

Topics

Resources

License

Contributing

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

`pocket_tts` (default)

`avspeech` — macOS built-in voices

`kokoro` — FluidAudio Kokoro

Packages