rhapsode-engine-chatterbox

Chatterbox as a rhapsode engine.

This is the engine the capability document was designed around, and the reason it has two levels. The turbo and nano builds perform the paralinguistic tags and discard the expressiveness dials; original and multilingual honour the dials and perform no tags. So on this engine you get cues or dials and never both, and which one is a fact about the weights that are resident right now.

nano is turbo's architecture with a smaller backbone, 110M parameters, loaded through turbo's own class. It claims exactly what turbo claims. It is in no upstream release yet, which is why this package installs upstream from a pinned commit on its master rather than from PyPI.

Its cues are performed, not only accepted. Its tokenizer has turbo's tags as the same tokens, and in September 2026 each of the eight lengthened the same cloned line on every one of three seeds, by more than on turbo ([chuckle] 1.0 s against 0.6 s). A level measure cannot tell a soft chuckle from a breath, so [chuckle] and [laugh] were checked by ear against turbo's, and both are there. One listen through the core alone had missed the chuckle, which is why it was checked twice.

From the root of a checkout, the SDK first, because the engine pins it exactly and neither is on PyPI yet:

python -m venv /opt/rhapsode/venvs/chatterbox
/opt/rhapsode/venvs/chatterbox/bin/pip install python/rhapsode-worker python/rhapsode-engine-chatterbox

pnpm wizard install chatterbox, or the web page, does this and the config below for you.

Then in rhapsode.config.json:

{ "engines": { "chatterbox": { "venv": "/opt/rhapsode/venvs/chatterbox", "env": { "CUDA_VISIBLE_DEVICES": "0" } } } }

The weights are MIT and commercially usable, which the catalog says before you install anything.

What it costs to install

Measured on an Apple Silicon Mac (MPS), September 2026:

• The virtualenv, torch included, is 1.5 GB.
• The weights come from Hugging Face the first time each variant loads, and land in ~/.cache/huggingface. turbo is 3.8 GB, and all three variants together are 9.7 GB.
• Loading turbo once it is cached takes about 10 s. A first /speak through the core (spawn, load, speak) took 17 s, and a warm one about 1 s for 2.4 s of audio.
• nano is a 3.0 GB fetch, of which it loads 1.9 GB: upstream's loader also downloads a 1.06 GB s3gen.safetensors that nano never reads.

Nano against turbo

Measured through the adapter on an Apple M5 Pro, September 2026, pinned upstream commit 5de7a54. One 98-character English line with a cue in it, about 5 to 6 s of audio. Each figure is the median of five warm runs after one discarded. "Realtime" is seconds of audio per second of wall clock.

build	device	load	stock voice	cloned voice
nano	MPS	9 s	5.0x realtime	3.7x realtime
turbo	MPS	13 s	2.7x realtime	1.3x realtime
nano	CPU	8–30 s	1.5–2.3x	0.9–1.4x
turbo	CPU	25 s	0.8x	0.9x

The CPU rows are ranges because the machine was busy while they were taken (a load average of 9 to 14 on 18 cores), and torch used 6 threads. Upstream says nano is 3x realtime on 8 CPU cores; it was not here. On a Mac, MPS is the device for either build.

The cloned column was measured when every request analysed its reference clip again, a second or more of each, because upstream's generate does that whenever it is handed a path. The adapter now analyses a voice once per resident build and keeps the result, so only a voice's first request after a load pays for it. Afterwards a warm clone cost what the stock voice did: nano 3.6x realtime against 3.6x, turbo 1.9x against 1.7x, taken once on a machine too busy for the absolute figures to stand beside the table's. An entry is 1.3 MB on nano and 0.8 MB on original, and the 32 most recently used voices are kept.

Apple Silicon works through MPS with no configuration. The residency figures in the protocol (memory stranded by an OOM, the share an unload reclaims) were measured on CUDA and are not verified on unified memory.