Request access to moore-tts-kokoro
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This model is part of a research benchmark on Mooré text-to-speech. Generated speech can imitate real voices, so access comes with conditions: respect the license (CC BY-NC 4.0), do not clone or imitate a real person's voice without their consent, say clearly that the audio is synthetic when you share it, and do not use the model for fraud, harassment, impersonation or political misinformation. Please read the card, its terms and licenses before requesting access, and make sure you understand them.
Log in or Sign Up to review the conditions and access this model content.
Mooré Kokoro-82M (fine-tuned)
hexgrad/Kokoro-82M fine-tuned on 54.6 hours of read Mooré, with a rule-based Mooré grapheme-to-phoneme front end. It speaks with fixed voices taken from the readers of the training recordings, and runs faster than real time on a CPU.
- CER 0.35 % on 150 held-out Bible verses (dev150, also used to choose this checkpoint), against 0.47 % for the real recordings
- MooreSpeechBench v1.0 test: CER 0.38 % against 0.54 % for the real recordings; test_unseen (new voices): 0.43 % against 0.73 %
- 13 fixed voices (12 readers of the training data, plus
Vmisc, which pools several minor readers), no cloning, 24 kHz output, real-time factor 0.48 on 4 threads of a laptop CPU (Windows, no GPU) - Part of a controlled benchmark of open TTS architectures on Mooré (same data, same splits, same evaluation)
Résumé en français. Kokoro-82M (hexgrad) affiné sur 54,6 heures de mooré lu, avec un module de conversion graphème-phonème pour le mooré fondé sur des règles. Il parle avec des voix fixes tirées des lecteurs des enregistrements d'entraînement et tourne plus vite que le temps réel sur un processeur. CER de 0,35 % sur 150 versets bibliques réservés au test, contre 0,47 % pour les vrais enregistrements ; 13 voix fixes (12 lecteurs des données d'entraînement, plus
Vmisc, qui regroupe plusieurs lecteurs secondaires), pas de clonage, sortie 24 kHz, facteur temps réel 0,48 sur 4 fils d'un processeur de portable (Windows, sans GPU). Il fait partie d'une comparaison contrôlée d'architectures de synthèse vocale ouvertes sur le mooré (mêmes données, mêmes partitions, même évaluation). Licence CC BY-NC 4.0 ; pour un usage commercial, écrivez à nyantudrealban@gmail.com. Accès sur demande. Sur MooreSpeechBench v1.0, son CER est de 0,38 % surtest(0,54 % pour les enregistrements réels) et de 0,43 % surtest_unseen, voix nouvelles (0,73 %).
Samples
Four verses drawn at random from the held-out dev set (chapters never seen in training). For each one: this model with the voicepack of the verse's reader, and the real recording of the same verse.
| Text | Voice | Type | Audio |
|---|---|---|---|
| Mam sã n zẽk m zugu, yãmb tãoosda maam wa gɩgemd sẽn digd bũmbu, la y le tar y pãngã n maand maam wẽnga. | V13 | this model (CER 0.0%) | |
| Mam sã n zẽk m zugu, yãmb tãoosda maam wa gɩgemd sẽn digd bũmbu, la y le tar y pãngã n maand maam wẽnga. | V13 | real recording of the same verse | |
| T'a Zeezi yeel-a: Mam sẽn gomd ne foomã, mam yaa yẽnda. | V04 | this model (CER 0.0%) | |
| T'a Zeezi yeel-a: Mam sẽn gomd ne foomã, mam yaa yẽnda. | V04 | real recording of the same verse | |
| A wilga tõnd t'a yãa malɛk sẽn wa n yaas a roogẽ wã, n yeele: 'Tʋm rap Zope tɩ b tɩ bool a Sɩmo b sẽn boond me t'a Pɩyɛɛrã, | V07 | this model (CER 1.7%) | |
| A wilga tõnd t'a yãa malɛk sẽn wa n yaas a roogẽ wã, n yeele: 'Tʋm rap Zope tɩ b tɩ bool a Sɩmo b sẽn boond me t'a Pɩyɛɛrã, | V07 | real recording of the same verse | |
| La sapɩlmã na n sarma kãgrã, la b na n sirga tẽngã fasɩ. | V11 | this model (CER 0.0%) | |
| La sapɩlmã na n sarma kãgrã, la b na n sirga tẽngã fasɩ. | V11 | real recording of the same verse |
Usage
# pip install kokoro==0.9.4 soundfile
# huggingface-cli download anyantudre/moore-tts-kokoro --local-dir moore-kokoro
python moore-kokoro/code/speak.py --voice V04 --text "Yãmb kɩtame tɩ nintãm lebg b rɩɩbo." --out out.wav
# from Python: code/speak.py has load(repo_dir) and speak(model, voicepack, text)
Training
| Base model | hexgrad/Kokoro-82M |
| Parameters | 82M (StyleTTS 2 architecture: PL-BERT text encoder, duration and pitch predictors, iSTFTNet decoder) |
| Output | 24000 Hz mono |
| Data | 54.6 h, 20,491 verses read by ~16 voices, jw.org Mooré Bible recordings (24 kHz), train chapters only; edge silences trimmed, loudness -20 LUFS |
- Rule-based Mooré G2P (
code/g2p_moore.py) that writes Mooré in symbols of Kokoro's existing phoneme set, so no embedding is added - Two-stage StyleTTS 2 fine-tuning after the kikiri-tts recipe: stage 1 (20,000 steps) trains the acoustic modules with the text aligner, stage 2 (15,000 steps) adds the duration and pitch predictors; SLM adversarial training off
- Batch 6, the largest that fits a 32 GB GPU at full precision
- One voicepack per reader, built from up to 64 of that reader's training clips
- One RTX 5090, 5.1 GPU-hours; snapshot chosen by CER of the eval ASR on dev150 (stage 2, step 11,000)
Results
| dev150 (150 verses, 20 held-out chapters, 9 voices) | CER | WER | CER on ɛ ɩ ʋ and nasals | Speaker sim. to prompt (ECAPA) | UTMOSv2 | F0 contour dist. (st) |
|---|---|---|---|---|---|---|
| Real recordings | 0.47 % | 1.85 % | 1.32 % | 0.73 | 3.03 | 0.00 |
| This model | 0.35 % | 1.55 % | 1.16 % | 0.74 | 3.17 | 1.43 |
CER: anyantudre/moore-asr-eval-w2vbert, greedy CTC, never trained on these chapters. Speaker similarity: cosine between ECAPA-TDNN (speechbrain/spkrec-ecapa-voxceleb) embeddings of the audio and of the voice prompt; for the real recordings it compares two different verses by the same reader. A model can score above that level when it copies the recording conditions of the prompt closely. UTMOSv2 is an English-trained naturalness predictor and is only a rough guide for Mooré. F0 contour distance: word-aligned DTW between synthesized and natural pitch, in semitones (lower is closer). Kokoro does not clone a voice: it speaks with the voicepack of the verse's reader, and its speaker similarity is measured against the same prompt clip as for the other systems of the benchmark.
MooreSpeechBench v1.0 test results
| Set | System | CER, judge 1 | ΔCER [95 % CI] | CER, judge 2 | ΔCER [95 % CI] | Speaker sim. (ECAPA) | UTMOSv2 |
|---|---|---|---|---|---|---|---|
test |
Real recordings | 0.54 % | 0.63 % | 0.712 | 2.97 | ||
test |
This model | 0.38 % | -0.16 [-0.24, -0.07] | 0.40 % | -0.24 [-0.34, -0.12] | 0.737 | 3.20 |
test_unseen |
Real recordings | 0.73 % | 0.90 % | 0.701 | 2.94 | ||
test_unseen |
This model | 0.43 % | -0.30 [-0.39, -0.21] | 0.44 % | -0.46 [-0.55, -0.38] | n/a (stand-in voice) | 3.29 |
test: 541 verses, 23 unseen chapters, voices heard in training.test_unseen: 836 verses, 40 chapters of new books, two new voices (V17, V18). Run once, with 3 synthesis seeds; CER is the mean over the seeds.- ΔCER is this model minus the real recordings of the same verses, in CER points, with a paired chapter
bootstrap (2,000 resamples). Judges: judge 1 at
b834419, judge 2 atfafc2ce; greedy CTC, moore-asr-norm 1.0. - For real recordings, speaker similarity compares two verses by the same reader, the ceiling for a model. UTMOSv2 is a diagnostic only: clean synthetic speech can score above the real recordings.
- dev150 above was also used to choose this checkpoint, so the test sets are the unbiased measure.
- This model cannot clone a voice. On
test_unseenit speaks with the training voice nearest in pitch (V17 as V11, V18 as V15), so its speaker similarity there is not meaningful.
Limitations
- Trained on read Bible speech: the prosody sounds like Bible reading, and other domains are untested.
- No voice cloning: the voices are the readers of the training recordings (
voices/). Kokoro has no reference-audio input. - The G2P is hand-written rules. Names and loanwords with unusual spellings can be read wrong, and characters it does not know are skipped.
- Tones are not written in Mooré orthography, so the model has to infer them from the text around each word. Apart from the F0 contour distance, nothing here measures tone.
License
CC BY-NC 4.0; the base model (hexgrad/Kokoro-82M) is Apache 2.0. For commercial use, contact nyantudrealban@gmail.com. The training audio comes from Bible recordings whose terms allow non-commercial use only and restrict redistribution; it is not included, apart from the short reference clips in samples/.
- Downloads last month
- 54