Voice

Talk to the agent and hear it answer — on-device speech recognition and synthesis, which engines run, and what leaves your machine.

New in v0.3.15, you can speak to the agent from the sidebar and, if you want, have its replies read back to you. Speech recognition and speech synthesis both run on this machine; no audio is sent to a speech service.

Get started

Open Settings → General and, under Speech Models, download Speech input. That alone is enough for dictation.
If you want spoken answers, also download a voice: Spoken replies (English), Chinese replies (Chinese and English), or Natural voice (20 languages, large). Then switch on Speak replies under Voice.
In the sidebar, click the microphone button next to the message box and talk. What you say becomes an ordinary message in the current conversation — same history, same tools, same execution tier.

The browser asks for microphone permission the first time. Click the button again (or the ✕ on the voice bar) to stop; voice input also stops by itself after a stretch of silence.

Interrupting. Start talking while a reply is being read out and it stops speaking, so you can redirect it without waiting.

Voice view (off by default) replaces the message bubbles with a waveform while voice is on. The replies are still generated, stored and searchable — only the display changes.

Speech input

Two recognition engines are routed by language:

Engine	Languages	Notes
SenseVoice (default)	Chinese (Mandarin), Cantonese, English, Japanese, Korean	Fast enough for live dictation; downloaded by Speech input
Whisper (`base` by default)	Other languages	Used only for languages SenseVoice cannot handle

Whisper is built in, but its weights are not part of the Speech Models downloader. To use it, put a Hugging Face–format Whisper model (config.json, tokenizer.json, model.safetensors — not whisper.cpp's ggml-*.bin) in <cache>/nevoflux/models/whisper-base/, or point [tts.whisper] model_path at it. Larger sizes are noticeably better outside English but cost memory: roughly 585 MB resident for base, 1.9 GB for small, 4.8 GB for large-v3-turbo (model alone).

The agent can also transcribe audio files you give it. Recordings longer than 30 seconds are split at pauses in speech before transcribing.

Spoken replies

Two synthesis engines, chosen automatically:

MOSS-TTS-Nano (Natural voice) — Chinese and 19 other languages, the better voice. Used whenever it is installed and fast enough on this machine.
Kokoro — the fallback. The Spoken replies download is English-only; the Chinese replies download is a newer Kokoro release that speaks Chinese and English.

"Fast enough" is measured, not guessed: the agent times real replies and keeps a median of recent measurements. If MOSS takes more than 0.85× real time to speak, the reply would stall, so Kokoro takes over. The Voice group tells you which engine is speaking and why; Re-measure clears the verdict so the next reply tries MOSS again (useful after closing a heavy program or updating a driver). The Voice list only offers voices of the engine that will actually speak.

Only the reply text is read: code blocks are skipped and Markdown marks are dropped. While replies are being spoken, the model is asked to write for the ear — lead with the conclusion in whole sentences and not let code, tables or lists carry the meaning. The answer still appears in full as text.

GPU. With Use GPU when it helps on (the default), speech synthesis may try a GPU — DirectML on Windows (any DX12 GPU: NVIDIA, AMD or Intel), Core ML on macOS — and keeps it only if it measures faster than the CPU here; it often is not. On Linux speech always runs on the CPU. If a GPU backend crashes or fails, the agent records it in [speech] failed_providers and does not try it again; delete the entry to retry (for example after a driver update). Turn the switch off to pin the CPU.

Privacy

Audio stays on this machine. Recognition and synthesis run locally; the speech models are only downloaded, never called.
The transcript is a normal message. Once recognized, what you said is sent to your active model provider exactly as typed text would be.
Downloads come from hf-mirror.com / Hugging Face and GitHub (some via the ghfast.top proxy). Every file is checked against a pinned size and SHA-256 digest, so a changed or substituted file is refused. Models are kept in <cache>/nevoflux/models/ — ~/.cache/ on Linux, ~/Library/Caches/ on macOS, %LOCALAPPDATA% on Windows.
ElevenLabs is separate: a cloud text-to-speech tool the agent can use only if you configure an ElevenLabs API key under [tts.elevenlabs]. Text sent to it leaves your machine.

Configuration

Speech preferences (Speak replies, Voice, GPU, Voice view) are browser settings. The engines are configured in config.toml; all fields are optional and the defaults look in the models folder above.

[tts.sensevoice]
# model_path  = "~/.cache/nevoflux/models/sensevoice-small.int8.onnx"
# tokens_path = "~/.cache/nevoflux/models/sensevoice-tokens.txt"

[tts.whisper]
default_size = "base"      # tiny | base | small | medium | large-v3-turbo
# model_path = "/path/to/whisper-small"   # directory in Hugging Face layout

[tts.kokoro]
# default_voice = "af_heart"

[tts.moss]
# enabled = false          # keep Kokoro even when MOSS is installed
# default_voice = "Junhao"

[speech]
# Written by the agent — edit only to troubleshoot.
# execution_provider = "cpu"   # auto (default) | cpu | directml | cuda | coreml
# rtf_budget = 0.85            # above this, MOSS is too slow and Kokoro speaks
# failed_providers = []        # backends that failed here; clear to retry