Request access to moore-tts-checkpoints
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 text-to-speech for Mooré. Requests are reviewed by hand, usually within a few days. Generated speech can imitate real voices, so access comes with a few conditions: respect the license (the licence of each folder: Apache-2.0, MIT, CC BY 4.0, CC BY-NC 4.0 or CC BY-NC-SA 4.0, see the card), 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.
Log in or Sign Up to review the conditions and access this model content.
Mooré TTS: training checkpoints and data curves
This repository holds the best checkpoint of every run in a data-budget study of text-to-speech for Mooré (Burkina Faso). Seven systems are trained on 1, 5 and 20 hours of read Mooré, and on the full 54.6 hours, with the same recipe and the same number of steps for every budget. The 1 h and 5 h subsets are drawn three times (seeds s0, s1, s2). Each run keeps the checkpoint with the lowest character error rate (CER) on dev150.
Training is still running. This card is rebuilt as runs finish: 31 of 56 runs archived (last update 2026-11-01).
| System | 1 h | 5 h | 20 h | 54.6 h (phase 1) |
|---|---|---|---|---|
| Kokoro-82M | 0.77 / 0.78 / 0.62 % | 0.48 / 0.51 / 0.49 % | 0.36 % | 0.34 % |
| OmniVoice (613M) | 3.39 / 1.82 / 2.09 % | 1.01 / 0.65 / 0.83 % | 0.65 % | 0.50 % |
| Spark-TTS (0.5B) | 11.37 % | 5.26 % | 2.03 % | 1.29 % |
| Chatterbox ML (LoRA) | 3.89 % | 2.26 % | 1.33 % | 1.50 % |
| F5-TTS (336M) | pending | pending | pending | 1.97 % |
| MMS-TTS (36M, all voices) | pending | pending | pending | pending |
| Pocket TTS teacher (24 layers) | 86.18 / 82.61 / 83.85 % | 54.06 / 54.35 / 53.49 % | 5.77 % | 1.44 % |
Several values in a cell are seeds s0 / s1 / s2.
CER is measured on dev150: 150 Bible verses from 20 held-out chapters, read by 9 speakers, transcribed by a wav2vec-BERT 2.0 recognizer trained for evaluation only (anyantudre/moore-asr-eval-w2vbert). The recognizer reaches 0.47 % on the real recordings, which is the floor for every number here. Lower is better.
Listen
Each table below has the same two held-out verses for every run, synthesized with the same settings (seed 0, the voice of the real reader). The real recordings:
| Text | Reader | Real recording | |
|---|---|---|---|
| Verse 1 | 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 | |
| Verse 2 | T'a Zeezi yeel-a: Mam sẽn gomd ne foomã, mam yaa yẽnda. | V04 |
Checkpoints by system
The folder of a run is <system>/<budget>_<seed>, for example kokoro/b5h_s1. It holds the best checkpoint,
the training config and the training log. "Best step" is the step of the kept checkpoint.
Kokoro-82M (kokoro/)
82M parameters (StyleTTS 2 architecture, iSTFTNet decoder), fine-tuned from hexgrad/Kokoro-82M in two stages with a rule-based Mooré grapheme-to-phoneme front end. One voicepack per reader. Licence of this folder: Apache-2.0. Released model: anyantudre/moore-tts-kokoro.
| Training data | Seed | Best step | CER | Verse 1 | Verse 2 | Weights |
|---|---|---|---|---|---|---|
| 0.95 h | s0 | 9000 | 0.77 % | kokoro/b1h_s0/snapshots/s2_0009000 (0.3 GB) |
||
| 0.93 h | s1 | 9000 | 0.78 % | kokoro/b1h_s1/snapshots/s2_0009000 (0.3 GB) |
||
| 0.93 h | s2 | 5000 | 0.62 % | kokoro/b1h_s2/snapshots/s2_0005000 (0.3 GB) |
||
| 4.61 h | s0 | 9000 | 0.48 % | kokoro/b5h_s0/snapshots/s2_0009000 (0.3 GB) |
||
| 4.62 h | s1 | 14000 | 0.51 % | kokoro/b5h_s1/snapshots/s2_0014000 (0.3 GB) |
||
| 4.61 h | s2 | 15000 | 0.49 % | kokoro/b5h_s2/snapshots/s2_0015000 (0.3 GB) |
||
| 18.38 h | s0 | 11000 | 0.36 % | kokoro/b20h_s0/snapshots/s2_0011000 (0.3 GB) |
||
| 54.6 h (all) | 11000 | 0.34 % | kokoro/b_all/snapshots/s2_0011000 (0.3 GB) |
OmniVoice (omnivoice/)
613M parameters, full fine-tune of OmniVoice. Zero-shot voice cloning from a reference clip. Licence of this folder: CC BY-NC 4.0. Released model: anyantudre/moore-tts-omnivoice.
| Training data | Seed | Best step | CER | Verse 1 | Verse 2 | Weights |
|---|---|---|---|---|---|---|
| 0.95 h | s0 | 2000 | 3.39 % | omnivoice/b1h_s0/snapshot-2000 (1.2 GB) |
||
| 0.93 h | s1 | 2000 | 1.82 % | omnivoice/b1h_s1/snapshot-2000 (1.2 GB) |
||
| 0.93 h | s2 | 2000 | 2.09 % | omnivoice/b1h_s2/snapshot-2000 (1.2 GB) |
||
| 4.61 h | s0 | 4000 | 1.01 % | omnivoice/b5h_s0/snapshot-4000 (1.2 GB) |
||
| 4.62 h | s1 | 2000 | 0.65 % | omnivoice/b5h_s1/snapshot-2000 (1.2 GB) |
||
| 4.61 h | s2 | 4000 | 0.83 % | omnivoice/b5h_s2/snapshot-4000 (1.2 GB) |
||
| 18.38 h | s0 | 6000 | 0.65 % | omnivoice/b20h_s0/snapshot-6000 (1.2 GB) |
||
| 54.6 h (all) | 8000 | 0.50 % | omnivoice/b_all/snapshot-8000 (1.2 GB) |
Spark-TTS (spark/)
Qwen2.5-0.5B language model over BiCodec tokens, full fine-tune of Spark-TTS-0.5B. Licence of this folder: CC BY-NC-SA 4.0. Released model: anyantudre/moore-tts-spark.
| Training data | Seed | Best step | CER | Verse 1 | Verse 2 | Weights |
|---|---|---|---|---|---|---|
| 0.95 h | s0 | 1000 | 11.37 % | spark/b1h_s0/snap_0001000 (1.0 GB) |
||
| 0.93 h | s1 | queued | ||||
| 0.93 h | s2 | queued | ||||
| 4.61 h | s0 | 3000 | 5.26 % | spark/b5h_s0/snap_0003000 (1.0 GB) |
||
| 4.62 h | s1 | queued | ||||
| 4.61 h | s2 | queued | ||||
| 18.38 h | s0 | 1000 | 2.03 % | spark/b20h_s0/snap_0001000 (1.0 GB) |
||
| 54.6 h (all) | 3000 | 1.29 % | spark/b_all/snap_0003000 (1.0 GB) |
Chatterbox ML (chatterbox/)
Rank-64 LoRA adapter on the T3 model of Chatterbox Multilingual; the folder holds the adapter only, the base model is loaded from ResembleAI/chatterbox. Licence of this folder: MIT. Released model: anyantudre/moore-tts-chatterbox.
| Training data | Seed | Best step | CER | Verse 1 | Verse 2 | Weights |
|---|---|---|---|---|---|---|
| 0.95 h | s0 | 5000 | 3.89 % | chatterbox/b1h_s0/snapshots/step_005000 (0.2 GB) |
||
| 0.93 h | s1 | queued | ||||
| 0.93 h | s2 | queued | ||||
| 4.61 h | s0 | 1000 | 2.26 % | chatterbox/b5h_s0/snapshots/step_001000 (0.2 GB) |
||
| 4.62 h | s1 | queued | ||||
| 4.61 h | s2 | queued | ||||
| 18.38 h | s0 | 1000 | 1.33 % | chatterbox/b20h_s0/snapshots/step_001000 (0.2 GB) |
||
| 54.6 h (all) | 4000 | 1.50 % | chatterbox/b_all/snapshots/step_004000 (0.2 GB) |
F5-TTS (f5/)
336M parameters, full fine-tune of F5-TTS v1 Base on Mooré characters. Licence of this folder: CC BY-NC 4.0. Released model: anyantudre/moore-tts-f5.
| Training data | Seed | Best step | CER | Verse 1 | Verse 2 | Weights |
|---|---|---|---|---|---|---|
| 0.95 h | s0 | queued | ||||
| 0.93 h | s1 | queued | ||||
| 0.93 h | s2 | queued | ||||
| 4.61 h | s0 | queued | ||||
| 4.62 h | s1 | queued | ||||
| 4.61 h | s2 | queued | ||||
| 18.38 h | s0 | queued | ||||
| 54.6 h (all) | 60000 | 1.97 % | f5/b_all/snapshots/upd_060000 (0.7 GB) |
MMS-TTS (mms/)
36M parameters, VITS. Fine-tuned from facebook/mms-tts-mos on all readers at once (multi-speaker). The released MMS model was trained on one reader only, so its 54.6 h point is a new run. Licence of this folder: CC BY-NC 4.0. Released model: anyantudre/moore-tts-mms-v04.
| Training data | Seed | Best step | CER | Verse 1 | Verse 2 | Weights |
|---|---|---|---|---|---|---|
| 0.95 h | s0 | queued | ||||
| 0.93 h | s1 | queued | ||||
| 0.93 h | s2 | queued | ||||
| 4.61 h | s0 | queued | ||||
| 4.62 h | s1 | queued | ||||
| 4.61 h | s2 | queued | ||||
| 18.38 h | s0 | queued | ||||
| 54.6 h (all) | queued |
Pocket TTS teacher (pocket/)
The 24-layer teacher of Pocket TTS (Kyutai), adapted to Mooré from its French checkpoint. The released Pocket model is a 6-layer student distilled from the 54.6 h teacher (folder pocket/student_b_all). Licence of this folder: CC BY 4.0. Released model: anyantudre/moore-tts-pocket.
| Training data | Seed | Best step | CER | Verse 1 | Verse 2 | Weights |
|---|---|---|---|---|---|---|
| 0.95 h | s0 | 6000 | 86.18 % | not kept (deleted before the archive existed; the score is from the registry) | ||
| 0.93 h | s1 | 2000 | 82.61 % | pocket/teacher_b1h_s1/checkpoint_00002000.pt (2.5 GB) |
||
| 0.93 h | s2 | 2000 | 83.85 % | pocket/teacher_b1h_s2/checkpoint_00002000.pt (2.5 GB) |
||
| 4.61 h | s0 | 2000 | 54.06 % | not kept (deleted before the archive existed; the score is from the registry) | ||
| 4.62 h | s1 | 2000 | 54.35 % | pocket/teacher_b5h_s1/checkpoint_00002000.pt (2.5 GB) |
||
| 4.61 h | s2 | 2000 | 53.49 % | pocket/teacher_b5h_s2/checkpoint_00002000.pt (2.5 GB) |
||
| 18.38 h | s0 | 4000 | 5.77 % | pocket/teacher_b20h_s0/checkpoint_00004000.pt (2.5 GB) |
||
| 54.6 h (all) | 6000 | 1.44 % | pocket/teacher_b_all/checkpoint_00006000.pt (2.5 GB) |
Other folders: eval_asr/w2vbert, pocket/ablation_en24l, pocket/ablation_en24l_cos8k, pocket/student_b_all. pocket/student_b_all is the distilled 6-layer Pocket
model that was released, pocket/ablation_* are the English-start ablations of the teacher, and eval_asr/w2vbert
is the recognizer used for every CER above.
Download one run
huggingface-cli download anyantudre/moore-tts-checkpoints --include "kokoro/b5h_s0/*" --local-dir moore-ckpt
Each family loads with the synth.py of its released model repo (linked in each section); pass the folder of
the checkpoint as --ckpt.
Data
All runs use the same training split of read Mooré: jw.org Bible recordings with their verse texts, 54.6 hours in total. The subsets are nested (1 h inside 5 h inside 20 h inside the full split), drawn chapter by chapter with the same mix of readers in each. Seeds s1 and s2 draw the 1 h and 5 h subsets again. dev150 comes from chapters that no run ever sees.
Licences
Each folder keeps the licence of its base model: Kokoro Apache-2.0, Chatterbox MIT, Pocket TTS CC BY 4.0, OmniVoice, F5-TTS and MMS CC BY-NC 4.0, Spark-TTS CC BY-NC-SA 4.0. The training audio is third-party Bible audio used for non-commercial research.
Notes
- Two Pocket teacher checkpoints (1 h and 5 h, seed s0) were deleted before this archive existed. Their scores stay in the tables; their weights are gone.
- The 54.6 h checkpoints of Kokoro, OmniVoice, Spark, Chatterbox and F5 are the released models.
- These checkpoints are research artefacts. Generated speech can imitate a real voice: say that audio is synthetic when you share it, and do not imitate a real person without their consent.