Turaco-nano-tts-wes
Turaco-nano-tts-wes is a lightweight text-to-speech model designed to generate spoken Cameroon Pidgin English (WES) from text.
This model is part of the Turaco family, an initiative exploring language and speech models for low-resource and underrepresented African languages. With a compact architecture and targeted training on Cameroon Pidgin recordings, Turaco Nano aims to make culturally relevant speech technology easier to run locally and integrate into everyday applications.
Turaco Nano explores a smaller alternative to the earlier Turaco CSM release, with 109.5 million parameters and 438.0 MB of FP32 weights, including the speech codec. Its weight file is about 7.5× smaller than the earlier CSM release. The deployment goal is practical CPU inference and a path toward edge devices, mobile applications and browser-based speech. The package verifies Python CPU inference and includes a Q8 WebAssembly browser conversion. Physical mobile-device speed and native-speaker quality benchmarks remain to be completed.
The model was trained using Common Voice Scripted Speech 27.0 — Cameroon Pidgin, available through Mozilla Data Collective.
This release is an initial, single-speaker research pilot. The repository provides the complete inference weights, tokenizer, a precomputed pilot voice, model configuration, training recipe and measured results. Pronunciation, accent and intonation still require native-speaker listening; the measured results and evaluation limits are documented below.
Measured results
| Measurement | Result |
|---|---|
| GPU | Tesla T4 |
| Optimizer steps | 200 |
| Training / validation / test clips | 897 / 50 / 51 |
| Baseline validation objective | 0.527815 |
| Final validation objective | 0.341064 |
| Relative objective decrease | 35.38% |
| Parameters, codec included | 109,501,826 |
| Weight-file bytes | 438,033,312 |
These are fixed-seed, teacher-forced objective measurements, not pronunciation accuracy or perceptual quality scores. All 51 test texts were generated by both the base and adapted models; no automatic English-ASR score is presented as Pidgin accuracy.
A smaller footprint for local speech
| Turaco release | Parameters | Weight file | Released weight precision |
|---|---|---|---|
| Turaco CSM | 1,632,100,161 | 3,297,957,042 bytes | FP16 |
| Turaco Nano | 109,501,826 | 438,033,312 bytes | FP32 |
Nano has about 14.9× fewer parameters. The weight-size comparison uses the actual released files and their listed precisions. Runtime dependencies, voice state and generated examples add storage; smaller weights alone do not establish mobile latency, memory use or speech-quality parity with CSM.
Data and training
Common Voice 27.0 Cameroon Pidgin: 998 selected clips, about 59 minutes, one anonymous speaker. Original transcript-grouped splits: 897 train / 50 validation / 51 test. Alignment exclusions and the actual accepted counts are in metrics/alignment_counts.json. Audio is mono 24 kHz. English CTC forced alignment is a proxy for Pidgin and may introduce boundary errors. The source corpus documentation declares CC0. Original Pidgin spelling is retained.
The pretrained tokenizer and acoustic sampler are retained; sampler weights are frozen.
Training uses AdamW, LR 1e-5, batch 1, accumulation 8,
FP32 on a T4, and EMA decay 0.95. Checkpoint: checkpoint_00000200.pt.
Exact upstream revision: a4a6054db281f0f34f3d44d626a972a44560dc83. Full recipe and progress are in metrics/.
Direct adaptation of a released student is an experimental alternative to upstream teacher
adaptation followed by depth distillation.
Detailed validation objectives
| Raw-weight objective | Before adaptation | Latest checkpoint |
|---|---|---|
| flow_loss | 0.439705 | 0.330442 |
| eos_loss | 0.881103 | 0.106222 |
| loss | 0.527815 | 0.341064 |
These are fixed-seed, teacher-forced objective measurements, not pronunciation accuracy.
The baseline and fine-tuned listening files are in samples/baseline/ and
samples/finetuned/, with matching text listed in each results.json.
Evaluation and limitations
Paired baseline/fine-tuned samples use identical text, training-only reference voice, temperature and seeds. Held-out objective losses are not pronunciation accuracy or MOS. Validation uses random prefix crops, and export uses EMA while logged validation uses raw weights. No automatic English-ASR WER is presented as Pidgin accuracy. Native-speaker listening is needed to judge words, accent and intonation. One speaker and read sentences do not establish unseen-speaker, long-form or conversational performance. No quality conclusion is implied by export success.
Use Turaco Nano
Weights: 438,033,312 bytes; loaded model parameters: 109,501,826.
The weight bytes include the codec; voice.safetensors stores the pilot voice state separately.
That state was computed with this exact model from a CC0 training recording, so users can
try the pilot voice without the source dataset. Recompute it after changing model weights.
CPU speed varies by device. Browser/mobile deployment requires a
compatible conversion, quantization if appropriate, and measured quality and runtime checks.
from pocket_tts import TTSModel
model = TTSModel.load_model(config="hf://fotiecodes/turaco-nano-tts-wes/wes.yaml")
state = model.get_state_for_audio_prompt("hf://fotiecodes/turaco-nano-tts-wes/voice.safetensors")
audio = model.generate_audio(state, "I di go for ma wok.")
Browser demo
The browser playground is provided as a standalone website. The demo runs this adapted model locally using the MIT-licensed Phonon WebAssembly runtime, version 0.7.1.
- Browser weights: 177,252,320 bytes (about 177 MB), plus about 2.6 MB of tokenizer, config and pilot voice; the WASM engine is a separate small download.
- GGUF v3 with 24 transformer matrices in Q8_0 and 147 FP32 tensors. The unused speaker encoder is omitted; the speech decoder is retained.
- Exact conversion source, settings and hashes: browser/manifest.json. Original FP32 release is unchanged.
- Dedicated CPU worker, standard WASM SIMD, original JSON tokenizer, precomputed pilot voice. Pidgin text is passed through without English normalization.
- Click Generate speech to download the voice on first use and speak automatically. Complete model files and the app are saved in browser storage for future offline visits. Closing the tab does not delete them; there is no application expiry or automatic model cleanup. Persistent-storage protection is requested when supported, but users or browsers can still clear site data. Download WAV is an optional explicit export.
- Desktop browser loading, actual audio generation and playback were verified. Successful generation does not establish perceptual quality or mobile performance. RAM use exceeds download size; speed varies by device.
License and attribution
Turaco Nano is developed by Clevaway. This release is fine-tuned from Kyutai Pocket TTS by Manu Orsini, Simon Rouard, Gabriel De Marmiesse, Václav Volhejn, Neil Zeghidour and Alexandre Défossez. The base weights are licensed under CC BY 4.0; upstream code is MIT. The weights were modified through fine-tuning on the dataset documented above. The browser runtime is Phonon, MIT.
Install the pinned source revision for reproduction; the inference package API is compatible with the export interface at that revision. Use voice data with permission and identify generated speech as synthetic. The original tokenizer, model config and reproducibility files preserve provenance.
- Downloads last month
- -
Model tree for fotiecodes/turaco-nano-tts-wes
Base model
kyutai/pocket-tts