Engines

Home · Manual index

Engines overview

An engine is a MODEL_SETUP key: bark, chatterbox, dia, f5, fish, higgs, kokoro, qwen, vibevoice, whisper, xtts, speecht5, parler, outetts, vits, edge, voxtral, voxcpm2, csm, orpheus. The GUI Setup cards, the Server spawn list (except whisper), tts.cmd tts <model>, and /api/tts/{model} all use these ids.

Whisper is an engine you install and load, but it does not speak. It transcribes for verification, SRT, and the Voices-tab Transcribe button.

Shared lifecycle (every speaking engine)

1. Install — Setup card Install, or tts.cmd install <id> --wait, or POST /api/setup/install/{id}. Wait until GET /api/setup/status reports ready (or packages_only for engines that fetch weights on first use, notably vits).
2. Check devices — tts.cmd devices / GET /api/devices. Compare free VRAM to MODEL_VRAM_ESTIMATE_GB plus the 1.5 GiB reserve.
3. Load — Server tab Spawn Worker, or tts.cmd model load <id> --device cuda:0, or POST /api/models/{id}/load. Poll until a worker is ready. Cold-load times vary from seconds to several minutes.
4. Inspect voices and schema — tts.cmd schema <id>, tts.cmd capabilities <id>, GET /api/tts/{id}/voices.
5. Optional dry run — tts.cmd dryrun <id> --text "...".
6. Speak — Testing tab Generate, or tts.cmd tts <id> --text "...", or POST /api/tts/{id} / /submit.
7. Unload — Server Kill, tts.cmd model unload <id>, POST /api/models/{id}/unload. Unload terminates the process group and runs targeted posix_fadvise on that model's files. It does not drop_caches for the whole VM.

Never load two huge engines at once on a single 24 GB card (qwen + voxtral will not fit with the reserve). Edge and VITS are the exceptions that do not need GPU.

How cloning vs built-in vs style works

MODEL_SETUP flags drive the Testing tab:

Flag	UI / API meaning
builtin_voices	Voice dropdown from /api/tts/{model}/voices
ref_audio	Saved Voice / Or Upload row
ref_text	Reference Text field (exact transcript of the clip)
style_tags	Insert style tag row

Pass a reference as:

• GUI Saved Voice → voice = filename in voices/
• GUI Upload → multipart reference_audio
• CLI --ref FILE → base64 plus reference_audio_name
• CLI/API --voice path-or-name
• VibeVoice also accepts JSON reference_audios (one to four paths in speaker order)

XTTS mode: "cloned" (default) vs "built-in" when voice is a named speaker such as Daisy Studious.

Planning table (VRAM estimate, timeout, chunk size, sample rate)

Estimates are not hard CUDA limits. Generation length, precision, and cloning inputs add transient allocations.

id	Display	VRAM est. GiB	Infer timeout s	Chunk chars	Sample rate	GPU?
edge	Edge TTS	0	120	3000	24000	Cloud, no GPU
vits	VITS	0	60	250	22050	No GPU needed
kokoro	Kokoro 82M	1	120	500	24000	CPU or GPU
speecht5	SpeechT5	1	120	250	16000	CPU or GPU
parler	Parler-TTS	2	90	300	44100	GPU preferred
f5	F5-TTS	3	600	250	24000	GPU
xtts	XTTS v2	3	600	250	24000	GPU
outetts	OuteTTS 1.0 0.6B	3	180	250	44100	GPU
chatterbox	Chatterbox	4	600	250	24000	GPU
csm	Sesame CSM-1B	5	300	300	24000	GPU
fish	Fish Speech	5	180	250	24000	GPU
bark	Bark	6	600	200	24000	GPU
vibevoice	VibeVoice	7	300	800	24000	GPU
dia	Dia 1.6B	8	180	400	44100	GPU
voxcpm2	VoxCPM2	8	300	400	48000	GPU
orpheus	Orpheus 3B	8	300	200	24000	GPU
whisper	Whisper	10 (large)	n/a	n/a	n/a	GPU for large
higgs	Higgs Audio 3B	16	180	500	24000	Large GPU; CPU listed on the Setup card
qwen	Qwen Omni 7B	16	180	500	24000	Large GPU
voxtral	Voxtral 4B TTS	20	300	600	24000	Very large GPU

Whisper size VRAM from config: tiny ~1GB, base ~1GB, small ~2GB, medium ~5GB, large ~10GB.

Override directories

Conflicting Python stacks live under /opt/tts_server/overrides/<name>. Shared coqui serves bark, xtts, and vits. Removing one of those three does not delete Coqui while another remains. kokoro, whisper, speecht5, edge, and csm use the base venv only.

Text chunking you should know about

The gateway splits long text before inference. Engine-aware rules:

• VibeVoice — keeps Speaker N: line boundaries; repeats the speaker prefix on fragments; max 800 chars.
• Dia — repeats the active [S1] / [S2] tag on each fragment; max 400 chars.
• VoxCPM2 — repeats a leading (voice design) on every fragment; max 400 chars.

If a take cuts off speaker two on Dia, the text was too long for one chunk's token budget; split turns yourself or rely on the bounded chunker.

Choosing an engine (usage, not marketing)

You want	Start with
Fast cloud, 300+ languages/locales, no install weight	edge
Fast local narration, many compact voices	kokoro
Lightweight CPU multi-speaker	speecht5
Fast CPU single English speaker	vits
Named speakers + multilingual clone	xtts
Two-person script with laughs	dia
Emotion slider on a cloned voice	chatterbox
Clone from reference + transcript (diffusion)	f5
Clone + emotion/tone markers, fast local	fish
Describe the voice in English prose	parler (or voxcpm2 parenthesized design)
Long-form / scene description / clone	higgs
Two built-in voices only (Chelsie / Ethan)	qwen
Multi-speaker Speaker N: narration	vibevoice
14-language compact clone	outetts
9-language presets + clone, non-commercial	voxtral
48 kHz design/clone, 30 languages	voxcpm2
Conversational continuation from a clip	csm
English character + <laugh> tags	orpheus
Expressive non-speech effects, history chaining	bark
Captions / verify speech	whisper

Read engines-catalog.md for install files, speak examples, clone rules, and failure notes per id.

Related pages

• Setup tab
• Server tab
• Testing tab
• Engines catalog
• A to Z knowledge page