Request access to moore-tts-spark

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-SA 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 Spark-TTS (fine-tuned)

SparkAudio/Spark-TTS-0.5B with its language model fully fine-tuned on 54.6 hours of read Mooré. It clones a voice from the BiCodec global tokens of a short reference clip.

  • CER 1.29 % on 150 held-out Bible verses (dev150, also used to choose this checkpoint), against 0.47 % for the real recordings and 58.12 % for the untouched base model
  • MooreSpeechBench v1.0 test: CER 1.55 % against 0.54 % for the real recordings; test_unseen (new voices): 1.50 % against 0.73 %
  • Zero-shot voice cloning from a reference clip, 16 kHz output, real-time factor 0.27 on one RTX 5090 (one verse at a time)
  • Part of a controlled benchmark of open TTS architectures on Mooré (same data, same splits, same evaluation)

Résumé en français. SparkAudio/Spark-TTS-0.5B dont le modèle de langue a été entièrement affiné sur 54,6 heures de mooré lu. Il clone une voix à partir des jetons globaux BiCodec d'un court extrait de référence. CER de 1,29 % sur 150 versets réservés, contre 0,47 % pour les vrais enregistrements et 58,12 % pour le modèle de base non modifié ; clonage de voix sans entraînement, sortie 16 kHz, facteur temps réel 0,27 sur une RTX 5090 (un verset à la fois). 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-SA 4.0, comme le modèle de base ; pour un usage commercial, écrivez à nyantudrealban@gmail.com. Accès sur demande. Sur MooreSpeechBench v1.0, son CER est de 1,55 % sur test (0,54 % pour les enregistrements réels) et de 1,50 % sur test_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, the untouched base model, and the reference clip of the voice being cloned.

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 SparkAudio/Spark-TTS-0.5B, no fine-tuning (CER 60.4%)
(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 0.0%)
T'a Zeezi yeel-a: Mam sẽn gomd ne foomã, mam yaa yẽnda. V04 SparkAudio/Spark-TTS-0.5B, no fine-tuning (CER 28.8%)
(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 SparkAudio/Spark-TTS-0.5B, no fine-tuning (CER 64.7%)
(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 SparkAudio/Spark-TTS-0.5B, no fine-tuning (CER 46.3%)
(voice prompt: a different verse by the same reader) V11 reference clip

Usage

git clone https://github.com/SparkAudio/Spark-TTS && cd Spark-TTS && pip install -r requirements.txt
huggingface-cli download anyantudre/moore-tts-spark --local-dir pretrained_models/moore-spark
python -m cli.inference --model_dir pretrained_models/moore-spark --save_dir out \
    --prompt_speech_path reference.wav --text "Yãmb kɩtame tɩ nintãm lebg b rɩɩbo."

Training

Base model SparkAudio/Spark-TTS-0.5B
Parameters 0.66B (Qwen2.5-0.5B LM 507M + BiCodec 156M)
Output 16000 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 Qwen2.5-0.5B language model; BiCodec and wav2vec2 kept frozen and used to precompute the speech tokens
  • Prompt layout identical to the upstream zero-shot prompt without prompt text (task token, text, 32 global tokens), so the official inference code works unchanged
  • 4,258 steps (8 epochs), lr 1e-4 with 200 warmup steps and cosine decay, about 24k tokens per step, bf16 autocast
  • One RTX 5090, 1.1 GPU-hours; snapshot chosen by CER of the eval ASR on dev150 (step 3,000)

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
SparkAudio/Spark-TTS-0.5B, no fine-tuning 58.12 % 94.54 % 128.41 % 0.44 2.96 2.53
This model 1.29 % 4.56 % 2.49 % 0.68 3.22 1.60

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, which is the ceiling. 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.

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 1.55 % +1.00 [+0.74, +1.30] 1.86 % +1.22 [+0.96, +1.51] 0.694 3.22
test_unseen Real recordings 0.73 % 0.90 % 0.701 2.94
test_unseen This model 1.50 % +0.77 [+0.61, +0.93] 1.81 % +0.91 [+0.73, +1.09] 0.612 3.22
  • 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 at fafc2ce; 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.

Limitations

  • Trained on read Bible speech: the prosody sounds like Bible reading, and other domains are untested.
  • Output is 16 kHz (BiCodec), lower bandwidth than the 24 kHz systems of the benchmark.
  • Do not pass --prompt_text: fine-tuning used the prompt without its transcript, and the continuation mode was not trained.

License

CC BY-NC-SA 4.0, inherited from SparkAudio/Spark-TTS-0.5B. 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
26
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for anyantudre/moore-tts-spark

Finetuned
(25)
this model

Space using anyantudre/moore-tts-spark 1