felt-voice

Felt's voice — local speech-to-text (faster-whisper) and text-to-speech (IndexTTS-2.5 with per-line moods, Chatterbox-Turbo or Piper, with Piper as the automatic fallback), all re-voiced through the same Kasane Teto RVC model. Part of the Felt project; see that repo for the overall architecture.

Both engines are local — no audio ever leaves the LAN, matching Felt's local-first principle. Config (never hardcoded) via .env — see .env.example.

Quickstart

uv sync
uv run python -m piper.download_voices --download-dir voices en_US-lessac-medium
# Chatterbox reference clip (not committed, models/ is ignored):
#   models/chatterbox/teto_ref.wav
cp .env.example .env
uv run felt-voice-serve

API

• POST /v1/tts — {"text": "...", "emotion": "sleepy"} → WAV audio bytes. emotion is optional (felt-core's mood words: happy, excited, amused, sad, annoyed, surprised, thinking, sleepy, neutral); only indextts uses it.
• POST /v1/stt — multipart file upload (audio) → {"text": "..."}.

TTS engines

indextts (live since 2026-09-29)

Mario's pick from the 2026-09-29 voice A/B: IndexTTS-2.5 reading with Teto's own (Japanese) reference clip, then the Teto RVC model at pitch -4 (FELT_VOICE_RVC_PITCH=-4). IndexTTS supplies natural delivery and the mood; RVC gives her the familiar timbre.

IndexTTS runs as a separate local service, felt-indextts (indextts_service/server.py, 127.0.0.1:8431), because its torch and transformers pins clash with this venv. It returns base audio; RVC stays in this process. Setup on the Nitro5:

deploy/install-indextts.sh                       # clone + venv + weights + reference
sudo cp deploy/felt-indextts.service /etc/systemd/system/
sudo systemctl daemon-reload && sudo systemctl enable --now felt-indextts

Measured on the RTX 3060 6 GB (2026-09-29): ~3.8 GB VRAM for IndexTTS + GPU RVC together (its 2.2 GB reference encoder runs on the CPU, once per reference), and a line takes about its own duration to synthesise (a 3 s first sentence ~3.7 s, a 10 s line ~10.5 s). The service warms up at start (the first call otherwise takes 13-23 s).

chatterbox / piper

FELT_VOICE_TTS_ENGINE=chatterbox|piper. Chatterbox-Turbo is loaded once at startup in low-VRAM mode (T3 in bf16, conditioning encoders on CPU, watermark off); Piper is loaded too, and takes over automatically, with a warning in the log, if Chatterbox fails to load or errors on a request. Every bracketed tag ([laugh], [sigh], ...) is stripped before synthesis unless FELT_VOICE_ALLOW_TAGS=true (see tags.py). Each request logs tts engine=... synth_s=... rvc_s=... total_s=....

Status

Built (2026-09-11): both endpoints verified end-to-end via a real TTS → STT round trip through the actual HTTP API, not just a script test. STT runs on CPU by default — the reference deployment (Nitro5, 6 GB VRAM) is already tight with Ollama loaded for felt-core; CPU int8 is fast enough for short utterances. Not yet built: mic capture (that's a client's job — the dashboard or overlay companion — not this service's), wake-word detection, RVC voice conversion.