caribbean-tts

An educational experiment fine-tuning Kyutai's released 24-layer English Pocket-TTS toward English spoken with a Trinidadian accent.

Model and training

The two checkpoints come from the same 24l_finetune_wrapup run, initialized from the released English english_2026-04_24l model. This is full-model fine-tuning, not a LoRA adapter or a model trained from scratch.

Fine-tuned on a small, cleaned collection of Trinidadian English speech as an educational accent-adaptation experiment. Training audio, transcripts, source details, and voice prompts are not included.

Training configuration: 8,000 steps, learning rate 2e-5, constant schedule with 500 warm-up steps, batch size 32, gradient accumulation 2, and EMA decay 0.999.

Checkpoints

File	Role
checkpoint_8000.pt	Final checkpoint; used for the later demos
checkpoint_6250.pt	Earlier comparison checkpoint from the same run

The unavailable 3,250- and 6,000-step checkpoints are not part of this release. “Final” identifies the completed run, not a measured superiority across all inputs.

Intended use and limitations

For educational exploration of Trinidadian English accent adaptation and comparative listening. This is an experimental model, not a production-ready speech system. No quantitative accent acceptance or broad naturalness claim is made.

Short phrases and unusual names can rush, slur, or be mispronounced. Users should check for silence, early endings, truncation, and runaway generation. The training set is small and unbalanced; it does not represent the full diversity of Trinidadian speech.

Voice conditioning affects the output voice. The project used mixed prompts for demos and does not intend to impersonate a source speaker. This does not guarantee anonymity. Supply your own authorized voice prompt; no training-derived prompt is distributed. Do not use generated speech to impersonate people deceptively.

Usage

Tested on Python 3.12 with Pocket-TTS 3.0.2, PyTorch 2.14.1, and SoundFile 0.14.0 on CPU. Both checkpoints loaded and produced finite, non-silent audio for a public-safe test sentence. This is a compatibility smoke test, not an accent-quality benchmark.

pip install pocket-tts==3.0.2 torch==2.14.1 soundfile==0.14.0 huggingface_hub

The matching 24-layer configuration is bundled with Pocket-TTS 3.0.2. The loader also retrieves upstream tokenizer and audio-codec assets; an internet connection is needed on first use.

from pathlib import Path
import pocket_tts
from pocket_tts import TTSModel
from huggingface_hub import hf_hub_download
import soundfile as sf

checkpoint = hf_hub_download("nash0x7e2/caribbean-tts", "checkpoint_8000.pt")
config = Path(pocket_tts.__file__).parent / "config" / "english_2026-04_24l.yaml"

model = TTSModel.load_model(
    config=str(config),
    checkpoint=checkpoint,
)
# Enable audio prompting for these fine-tuned checkpoints.
model.has_voice_cloning = True
state = model.get_state_for_audio_prompt("your_authorized_prompt.wav")
audio = model.generate_audio(state, "Hello there. How are you today?")
sf.write("output.wav", audio.detach().cpu().numpy(), model.sample_rate)

Select checkpoint_6250.pt in hf_hub_download for the earlier comparison. Use your own authorized prompt WAV; no voice prompt is bundled.

Attribution

Adapted from Kyutai Pocket-TTS, base revision 492522650173a0653b7575cdc25ae09810e5d741, released under CC BY 4.0. This release modifies the base weights through Trinidadian English accent fine-tuning. Kyutai does not endorse this adaptation.

Code

Data preparation and evaluation tools: caribbean-tts-experiment on GitHub.