Request access to moore-tts-f5

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 (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.

Mooré F5-TTS (fine-tuned)

F5-TTS v1 Base fully fine-tuned on 54.6 hours of read Mooré. It is a non-autoregressive flow-matching model that clones a voice from a short reference clip and its transcript.

CER 1.97 % on 150 held-out Bible verses, against 0.47 % for the real recordings and 28.83 % for the base model with its vocabulary extended
Zero-shot voice cloning from a reference clip, 24 kHz output, real-time factor 0.06 on one RTX 5090 (32 flow steps, fp16)
Part of a controlled benchmark of open TTS architectures on Mooré (same data, same splits, same evaluation)

Samples

Four verses drawn at random from the held-out dev set (chapters never seen in training). For each one: this model, the base model with its vocabulary extended, and the reference clip of the voice being cloned.

Text	Voice	Type
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 2.0%)
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	F5-TTS v1 Base, no Mooré training (vocabulary extended only) (CER 20.8%)
(voice prompt: a different verse by the same reader)	V13	reference clip
T'a Zeezi yeel-a: Mam sẽn gomd ne foomã, mam yaa yẽnda.	V04	this model (CER 1.9%)
T'a Zeezi yeel-a: Mam sẽn gomd ne foomã, mam yaa yẽnda.	V04	F5-TTS v1 Base, no Mooré training (vocabulary extended only) (CER 13.5%)
(voice prompt: a different verse by the same reader)	V04	reference clip
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	F5-TTS v1 Base, no Mooré training (vocabulary extended only) (CER 35.3%)
(voice prompt: a different verse by the same reader)	V07	reference clip
La sapɩlmã na n sarma kãgrã, la b na n sirga tẽngã fasɩ.	V11	this model (CER 1.9%)
La sapɩlmã na n sarma kãgrã, la b na n sirga tẽngã fasɩ.	V11	F5-TTS v1 Base, no Mooré training (vocabulary extended only) (CER 35.2%)
(voice prompt: a different verse by the same reader)	V11	reference clip

Usage

# pip install f5-tts==1.1.22
# huggingface-cli download anyantudre/moore-tts-f5 --local-dir moore-f5
# lines.jsonl: {"id": "1", "text": "Yãmb kɩtame tɩ nintãm lebg b rɩɩbo.", "voice": "x",
#               "prompt_path": "reference.wav", "prompt_text": "Transcript of reference.wav"}
python moore-f5/code/synth.py --ckpt moore-f5 --manifest lines.jsonl --out_dir out --device cuda

Training


Base model	SWivid/F5-TTS
Parameters	336M (F5-TTS v1 Base DiT) + 13.5M Vocos vocoder
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

Full fine-tuning of the text encoder and the DiT, starting from F5TTS_v1_Base (1.25M updates)
Character vocabulary extended with the Mooré letters; each new embedding row copied from its base letter (ɛ from e, ɩ from i, ʋ from u, nasal vowels from the plain vowel)
Text fed as characters, without the upstream pinyin conversion
60,000 updates, AdamW lr 1e-5, 1,000 warmup updates then linear decay to 0, bf16, about 128 s of audio per update
One RTX 5090, 4.2 GPU-hours; snapshot chosen by CER of the eval ASR on dev150 (the last one, 60,000: CER kept falling until the learning rate reached 0)

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
F5-TTS v1 Base, no Mooré training (vocabulary extended only)	28.83 %	66.97 %	84.60 %	0.76	3.05	2.45
This model	1.97 %	7.14 %	12.12 %	0.80	2.91	1.86

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). Every system clones the same reference clip per verse.

Limitations

Trained on read Bible speech: the prosody sounds like Bible reading, and other domains are untested.
The reference clip needs its transcript (at most 15 s of audio). Text is read character by character: use code/synth.py, since the stock F5-TTS command line converts text to pinyin first and refuses nothing.
The duration of each sentence is estimated from its length in bytes, as upstream does, so very short or very long sentences can sound rushed or slow.
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, inherited from SWivid/F5-TTS. The training audio comes from Bible recordings whose terms restrict redistribution; it is not included, apart from the short reference clips in samples/.

Downloads last month: -

Safetensors

Model size

0.3B params

Tensor type

F16

Model tree for anyantudre/moore-tts-f5

Base model

SWivid/F5-TTS

Finetuned

(154)

this model