DevVoice

DevVoice is a desktop app and local HTTP API for speaking selected text with several local TTS models, including VibeVoice Realtime 0.5B and the higher-quality VibeVoice 1.5B model.

VibeVoice 1.5B usage now

The new 1.5B flow keeps the model practical by using a persistent MLX worker on macOS when:

• the selected model is VibeVoice
• precision is auto
• the MLX Python environment exists under ~/Library/Application Support/com.isa.devvoice/vibevoice-mlx-venv

If the MLX environment is missing, DevVoice now provisions it automatically on first 1.5B use. The old non-accelerated Rust 1.5B fallback is no longer treated as a supported path.

Do I need to warm up the model manually?

No, not for normal use.

When the app opens, DevVoice already starts a background warmup for the currently selected model. For VibeVoice 1.5B with MLX available, that startup warmup loads the MLX model and runs a tiny inference pass automatically.

The manual warmup API is still useful, but now it is optional and mainly for reducing the first request time for a specific voice.

Recommended user flow after opening the app

1. Open DevVoice.
2. Select VibeVoice as the voice model.
3. Leave precision set to auto.
4. Wait until the UI says the model is ready, or check GET /status and wait for "model_ready": true.
5. If you already have a preset, call /speak with reference_preset_name or reference_preset_id.
6. If you want the first request for that voice to be faster, call /vibevoice/warmup once with that preset before the first real /speak.
7. Reuse the same preset in the same app session for the best repeat performance.

Status check

You can confirm the active backend with:

curl http://127.0.0.1:9876/status

When the MLX path is active for 1.5B, tts_runtime_label should show:

{
  "model_ready": true,
  "tts_runtime_label": "python:mlx-int8:no-semantic"
}

MLX setup for the accelerated 1.5B path

DevVoice now auto-provisions the MLX runtime on first 1.5B use. If you want to install it yourself ahead of time, you can still create the virtual environment manually:

DATA_DIR="$HOME/Library/Application Support/com.isa.devvoice"
python3 -m venv "$DATA_DIR/vibevoice-mlx-venv"
source "$DATA_DIR/vibevoice-mlx-venv/bin/activate"
pip install --upgrade pip
pip install 'git+https://github.com/gafiatulin/vibevoice-mlx.git@f513aa7877e77fefa1aebe87432855c407da3b87' scipy

After that, relaunch DevVoice and keep precision on auto for the MLX worker to be used.

API calls

Create a preset once

curl -X POST "http://127.0.0.1:9876/vibevoice/presets" \
  -H "Content-Type: application/json" \
  -d '{"name":"my-demo-voice","referenceAudioPath":"/Users/you/Desktop/reference-voice.wav"}'

Names are unique, so reference_preset_name is safe to use later. New presets now copy the reference clip into DevVoice-managed storage so they do not depend on the original temporary source path.

List presets

curl http://127.0.0.1:9876/vibevoice/presets

Optional preset warmup

curl -X POST "http://127.0.0.1:9876/vibevoice/warmup" \
  -H "Content-Type: application/json" \
  -d '{"referencePresetName":"my-demo-voice"}'

Use this when you want the first real synthesis for that voice to avoid paying the model-load and voice-encode cost right before playback.

Speak with a preset by name

curl -X POST "http://127.0.0.1:9876/speak?reference_preset_name=my-demo-voice&cfg_scale=1.3&temperature=0.0&max_tokens=96&save_audio=true" \
  -d "I'm your worst nightmare."

Speak with a preset by id

curl -X POST "http://127.0.0.1:9876/speak?reference_preset_id=my-demo-voice-1716420000&cfg_scale=1.3&temperature=0.0&max_tokens=96&save_audio=true" \
  -d "I'm your worst nightmare."

Speak directly from a raw reference audio path

curl -X POST "http://127.0.0.1:9876/speak?reference_audio_path=/Users/you/Desktop/reference-voice.wav&cfg_scale=1.3&temperature=0.0&max_tokens=96&save_audio=true" \
  -d "I'm your worst nightmare."

Preset durability note

New presets now keep a DevVoice-managed copy of the reference clip, so they no longer depend on the original source path remaining available. Older presets created before this change may still point at external files.

Practical latency expectations

With the persistent MLX path working, the rough behavior on the tested machine was:

• startup warmup in about 20 seconds
• first preset-backed request in about 30 seconds
• repeated request with the same preset in the same app session in about 6 seconds

That is still slower than VibeVoice Realtime 0.5B, but it is dramatically faster than the previous tens-of-minutes Rust path for VibeVoice 1.5B.

More detail

See architecture.md for the current 1.5B request lifecycle, component layout, fallback behavior, and known limitations.