Chatterbox Nano, long utterances (British English)

A fine-tune of the T3 stage of ResembleAI/chatterbox-nano that carries a long sentence across several passes. Nano speaks about 24 words in one pass, so a longer sentence has to be cut into chunks, and the stock model starts every chunk as if it were a new sentence. This model was trained to continue instead: it is told where a chunk sits in its sentence and is given the previous chunk as already spoken.

It was trained on British English narration. This is round m7.

Contents

File Precision Size Use
t3_prefill_ane.mlpackage fp16 180 MB Prefill half of T3
t3_decode_ane.mlpackage fp16 180 MB Decode half of T3
pytorch/t3_long_utterance_m7.safetensors fp32 873 MB The merged T3 weights the packages were exported from

Everything else comes unchanged from seanll95/chatterbox-nano-coreml: the S3Gen flow and vocoder, t3_embeddings.bin and the tokenizer. The two packages here replace the two of the same name there and use the same split, cache length (896) and inputs.

No voice is included

This repository holds no speaker conditionals (nano_conds.bin). The fine-tune changed the speaker projection inside T3, so the conditioning rows must be computed with these weights from your own reference audio; the safetensors file is here for that. The stock nano_conds.bin was computed with the stock projection and is not a match.

How to drive it

A unit of 24 words or fewer is one chunk and is used exactly like stock Nano. Only a single sentence longer than 24 words is cut, into roughly equal chunks, preferring a cut after , ; : β€” – ) ].

Markers. Three token ids are added around a chunk's text ids after tokenising:

Id Where Meaning
217 first id this text starts mid-sentence
218 last id the sentence goes on; this chunk ends at punctuation
219 last id the sentence goes on; forced cut, no pause

For chunk j of n (0-based):

  • j == 0: text is chunk 0. Ids are tokens(text) + [218 or 219].
  • j >= 1: text is chunk j-1, a space, then chunk j. Ids are [217] (only when j-1 > 0) + tokens(text) + the right marker for chunk j (only when j < n-1).

Capitalise the first letter only when the text starts the sentence, and add a closing full stop only when the chunk ends it.

Continuation. For j >= 1, prefill these rows, each embedding + wpe[position]:

Rows Content
0-375 376 conditioning rows
next T the text ids above
next 1 start-of-speech, 6561
next P chunk j-1's speech tokens as generated, without 6561 or 6562

Read logits from the last of those rows and decode from the next position. If fewer than 260 positions would be left of the 896, drop the prefix: text is chunk j alone, ids are [217] + tokens + right marker, and P is 0.

Sampling. Always sample: temperature 0.5, top-p 0.9, min-p 0.05, repetition penalty 1.2. The penalty history starts as {6561} for every chunk and takes only that chunk's new tokens.

Joining audio. Vocode each later chunk with up to 25 of the previous chunk's speech tokens in front, discard that lead-in, and crossfade 10 ms (240 samples at 24 kHz) into the previous chunk's tail. Insert no silence inside a sentence.

Measured

The yardstick is a larger Chatterbox model tuned on the same narration ("the teacher"). Each figure is the share of this model's renders that sit as close to the teacher's renders as the teacher's own reruns sit to each other (its 95th percentile). All were measured in PyTorch with the sampling above, before the Core ML export.

Test Spectral Pitch shape Pitch level Pace Join pitch Join pace
24 long sentences, 4 seeds 96 94 95 95 92 89
24 sentences of 6-24 words, 4 seeds 97-99 92-95 92-97 96-99

One render set moves a figure by about 5 points. Not measured: units of 17-24 words made of several packed sentences, and sentences that need four or more chunks.

No watermarking

Upstream Chatterbox applies Resemble AI's PerTh watermarker to every generated file. This repository holds T3 only, and audio made with it is not watermarked. Please use it responsibly.

Licence and attribution

MIT, inherited from the upstream model. Copyright Β© 2025 Resemble AI; see LICENSE.

@misc{chatterboxtts2025,
  author       = {{Resemble AI}},
  title        = {{Chatterbox-TTS}},
  year         = {2025},
  howpublished = {\url{https://github.com/resemble-ai/chatterbox}},
  note         = {GitHub repository}
}
Downloads last month
6
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for seanll95/chatterbox-nano-long-utterance-en_gb

Quantized
(18)
this model