Update README.md

0f803ce verified 6 days ago

25.2 kB

	---
	language:
	- nnh
	license: cc-by-nc-sa-4.0
	pipeline_tag: text-to-speech
	tags:
	- text-to-speech
	- tts
	- voice-cloning
	- zero-shot
	- malagasy
	- nnh
	- low-resource
	- onnx
	- bluetts
	base_model: notmax123/blue-v2
	datasets:
	- mimba/multivoice-nnh
	---

	# BlueTTS — Ngiemboon (`plt`), clonage zero-shot preserve

	Finetune de [BlueTTS](https://github.com/maxmelichov/BlueTTS) (base `notmax123/blue-v2`)
	pour parler ngiemboon (Plateau Ngiemboon, `plt`) sans perdre la capacite de
	clonage vocal zero-shot du modele de base.

	Chaque checkpoint est publie avec son bundle ONNX (inference CPU), ses **audios
	d'evaluation et ses metriques** — rien n'est efface, tu peux choisir le tien.

	## Le probleme resolu

	Finetuner un TTS de clonage sur peu de locuteurs detruit l'espace locuteur : le modele
	oublie le clonage et repete des syllabes. Un run precedent sur 4 voix l'a montre
	sans ambiguite : `sim_ecapa` s'effondrait de 0.45 a 0.22 en 1 000 steps.

	La parade retenue : entrainer sur 524 voix distinctes parlant un ngiemboon a accent natif
	([`mimba/multivoice-nnh`](https://huggingface.co/datasets/mimba/multivoice-nnh), 26 200 clips),
	pour que le modele apprenne a separer le contenu du locuteur. La diversite remplace le gel.

	## Evolution

	![evolution](assets/evolution.png)

	\| step \| loss \| sim moy \| min \| max \| vues \| inedites \|
	\|---:\|---:\|---:\|---:\|---:\|---:\|---:\|
	\| 0 \| 0.282 \| 0.324 \| 0.111 \| 0.523 \| 0.407 \| 0.282 \|
	\| 1000 \| 0.234 \| 0.414 \| 0.257 \| 0.654 \| 0.436 \| 0.392 \|
	\| 2000 \| 0.230 \| 0.408 \| 0.255 \| 0.585 \| 0.431 \| 0.385 \|
	\| 3000 \| 0.228 \| 0.405 \| 0.257 \| 0.589 \| 0.426 \| 0.383 \|
	\| 4000 \| 0.227 \| 0.388 \| 0.258 \| 0.528 \| 0.402 \| 0.374 \|
	\| 5000 \| 0.226 \| 0.391 \| 0.246 \| 0.577 \| 0.412 \| 0.369 \|
	\| 6000 \| 0.224 \| 0.369 \| 0.164 \| 0.604 \| 0.382 \| 0.355 \|
	\| 7000 \| 0.224 \| 0.348 \| 0.157 \| 0.513 \| 0.368 \| 0.328 \|
	\| 8000 \| 0.223 \| 0.357 \| 0.186 \| 0.507 \| 0.359 \| 0.354 \|
	\| 9000 \| 0.222 \| 0.328 \| 0.085 \| 0.458 \| 0.336 \| 0.321 \|

	- sim moy — similarite locuteur ECAPA-TDNN (`speechbrain/spkrec-ecapa-voxceleb`),
	cosinus entre l'audio genere et l'audio de reference, moyenne sur 4 references x N phrases.
	- vues / inedites — phrases d'evaluation initiales vs phrases jamais vues a
	l'entrainement (mesure de generalisation).

	Au step 9000, les phrases inedites obtiennent 0.321 contre 0.336 pour les phrases d'evaluation initiales : l'ecart (-0.015) montre que le modele generalise et ne recite pas son corpus.

	\| Checkpoint \| Voix1_p1 \| Voix1_p2 \| Voix1_p3 \| Voix1_p4 \| Voix2_p1 \| Voix2_p2 \| Voix2_p3 \| Voix2_p4 \| Voix3_p1 \| Voix3_p2 \| Voix3_p3 \| Voix3_p4 \| Voix4_p1 \| Voix4_p2 \| Voix4_p3 \| Voix4_p4 \|
	\| :------------ \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \|
	\| step_1000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix4_p4.wav"></audio> \|
	\| step_2000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix4_p4.wav"></audio> \|
	\| step_3000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix4_p4.wav"></audio> \|
	\| step_4000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix4_p4.wav"></audio> \|
	\| step_5000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix4_p4.wav"></audio> \|
	\| step_6000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix4_p4.wav"></audio> \|
	\| step_7000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix4_p4.wav"></audio> \|
	\| step_8000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix4_p4.wav"></audio> \|
	\| step_9000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix4_p4.wav"></audio> \|


	## Recette d'entrainement

	\| \| \|
	\|---\|---\|
	\| Base \| `notmax123/blue-v2` (AE `blue_codec.safetensors`) \|
	\| Dataset \| `mimba/multivoice-nnh` — 26 197 clips, 524 locuteurs (50 chacun) \|
	\| Texte \| IPA pre-phonemise (`text_phonemized`), `plt` \|
	\| Duration Predictor \| reentraine de zero sur les 524 voix, 15 000 steps (loss 1.3 -> 0.059) \|
	\| TTL \| finetune, batch 7 x accumulation 8 (batch effectif 56) \|
	\| LR \| 2.5e-4, reduit a 1.25e-4 apres plateau de loss + baisse de `sim_ecapa` \|
	\| Checkpoints \| tous les 1 000 steps, exportes en ONNX et evalues automatiquement \|

	Pourquoi le DP est reentraine : c'est lui qui portait le begaiement du run 4-voix
	(il est conditionne par le locuteur). Sur 524 voix, il apprend un rythme robuste.

	## Utilisation (ONNX, CPU)

	```python
	from huggingface_hub import snapshot_download
	d = snapshot_download("mimba/bluetts-nnh", allow_patterns=["eval/step_9000/onnx/*"])
	# -> text_encoder / vector_estimator / vocoder / duration_predictor
	# + codec_encoder / style_encoder / duration_style_encoder (clonage zero-shot)
	```

	Parametres de production (worker Mimba) : `total_step=8..16`, `speed=0.95`,
	`cfg_scale=4.0`, `pace_blend=0.30`. Normaliser la sortie avant ecriture
	(RMS 0.08 / peak 0.95) : le vocodeur sort a un pic ~1.5 et s'ecrete sinon.

	Le vocabulaire BlueTTS ne contient pas les marqueurs de prenasalisation ngiemboons
	`m` en exposant et `n` en exposant : les remplacer par `m` / `n` avant encodage.

	## Limites

	- Fidelite de clonage inferieure a la base : `sim_ecapa` passe de ~0.35 (step 0)
	a ~0.32 apres finetune. La degradation a ete stoppee (LR reduit), pas annulee.
	- Variable selon la voix : certaines references se clonent nettement moins bien
	(voir l'ecart min/max du tableau) — tester avec ses propres references.
	- Monolingue : entraine pour le ngiemboon uniquement.
	- Donnees synthetiques : le corpus multi-voix est genere (OmniVoice + OpenVoice),
	seul l'accent provient de locuteurs natifs reels.

	## Licence

	`cc-by-nc-sa-4.0` — usage non commercial. Verifier independamment les licences de
	BlueTTS, OmniVoice et OpenVoice avant tout usage commercial.

	##### Contact : [@Mimba](baounabaouna@gmail.com)

	---
	language:
	- nnh
	license: cc-by-nc-sa-4.0
	pipeline_tag: text-to-speech
	tags:
	- text-to-speech
	- tts
	- voice-cloning
	- zero-shot
	- malagasy
	- nnh
	- low-resource
	- onnx
	- bluetts
	base_model: notmax123/blue-v2
	datasets:
	- mimba/multivoice-nnh
	---

	# BlueTTS — Ngiemboon (`plt`), clonage zero-shot preserve

	Finetune de [BlueTTS](https://github.com/maxmelichov/BlueTTS) (base `notmax123/blue-v2`)
	pour parler ngiemboon (Plateau Ngiemboon, `plt`) sans perdre la capacite de
	clonage vocal zero-shot du modele de base.

	Chaque checkpoint est publie avec son bundle ONNX (inference CPU), ses **audios
	d'evaluation et ses metriques** — rien n'est efface, tu peux choisir le tien.

	## Le probleme resolu

	Finetuner un TTS de clonage sur peu de locuteurs detruit l'espace locuteur : le modele
	oublie le clonage et repete des syllabes. Un run precedent sur 4 voix l'a montre
	sans ambiguite : `sim_ecapa` s'effondrait de 0.45 a 0.22 en 1 000 steps.

	La parade retenue : entrainer sur 524 voix distinctes parlant un ngiemboon a accent natif
	([`mimba/multivoice-nnh`](https://huggingface.co/datasets/mimba/multivoice-nnh), 26 200 clips),
	pour que le modele apprenne a separer le contenu du locuteur. La diversite remplace le gel.

	## Evolution

	![evolution](assets/evolution.png)

	\| step \| loss \| sim moy \| min \| max \| vues \| inedites \|
	\|---:\|---:\|---:\|---:\|---:\|---:\|---:\|
	\| 0 \| 0.282 \| 0.324 \| 0.111 \| 0.523 \| 0.407 \| 0.282 \|
	\| 1000 \| 0.234 \| 0.414 \| 0.257 \| 0.654 \| 0.436 \| 0.392 \|
	\| 2000 \| 0.230 \| 0.408 \| 0.255 \| 0.585 \| 0.431 \| 0.385 \|
	\| 3000 \| 0.228 \| 0.405 \| 0.257 \| 0.589 \| 0.426 \| 0.383 \|
	\| 4000 \| 0.227 \| 0.388 \| 0.258 \| 0.528 \| 0.402 \| 0.374 \|
	\| 5000 \| 0.226 \| 0.391 \| 0.246 \| 0.577 \| 0.412 \| 0.369 \|
	\| 6000 \| 0.224 \| 0.369 \| 0.164 \| 0.604 \| 0.382 \| 0.355 \|
	\| 7000 \| 0.224 \| 0.348 \| 0.157 \| 0.513 \| 0.368 \| 0.328 \|
	\| 8000 \| 0.223 \| 0.357 \| 0.186 \| 0.507 \| 0.359 \| 0.354 \|
	\| 9000 \| 0.222 \| 0.328 \| 0.085 \| 0.458 \| 0.336 \| 0.321 \|

	- sim moy — similarite locuteur ECAPA-TDNN (`speechbrain/spkrec-ecapa-voxceleb`),
	cosinus entre l'audio genere et l'audio de reference, moyenne sur 4 references x N phrases.
	- vues / inedites — phrases d'evaluation initiales vs phrases jamais vues a
	l'entrainement (mesure de generalisation).

	Au step 9000, les phrases inedites obtiennent 0.321 contre 0.336 pour les phrases d'evaluation initiales : l'ecart (-0.015) montre que le modele generalise et ne recite pas son corpus.

	\| Checkpoint \| Voix1_p1 \| Voix1_p2 \| Voix1_p3 \| Voix1_p4 \| Voix2_p1 \| Voix2_p2 \| Voix2_p3 \| Voix2_p4 \| Voix3_p1 \| Voix3_p2 \| Voix3_p3 \| Voix3_p4 \| Voix4_p1 \| Voix4_p2 \| Voix4_p3 \| Voix4_p4 \|
	\| :------------ \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \| :--------------------------------------------------------------------------------------------------------------- \|
	\| step_1000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_1000/voix4_p4.wav"></audio> \|
	\| step_2000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_2000/voix4_p4.wav"></audio> \|
	\| step_3000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_3000/voix4_p4.wav"></audio> \|
	\| step_4000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_4000/voix4_p4.wav"></audio> \|
	\| step_5000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_5000/voix4_p4.wav"></audio> \|
	\| step_6000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_6000/voix4_p4.wav"></audio> \|
	\| step_7000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_7000/voix4_p4.wav"></audio> \|
	\| step_8000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_8000/voix4_p4.wav"></audio> \|
	\| step_9000 \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix1_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix1_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix1_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix1_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix2_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix2_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix2_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix2_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix3_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix3_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix3_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix3_p4.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix4_p1.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix4_p2.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix4_p3.wav"></audio> \| <audio controls src="https://huggingface.co/mimba/bluetts-nnh/resolve/main/eval/step_9000/voix4_p4.wav"></audio> \|


	## Recette d'entrainement

	\| \| \|
	\|---\|---\|
	\| Base \| `notmax123/blue-v2` (AE `blue_codec.safetensors`) \|
	\| Dataset \| `mimba/multivoice-nnh` — 26 197 clips, 524 locuteurs (50 chacun) \|
	\| Texte \| IPA pre-phonemise (`text_phonemized`), `plt` \|
	\| Duration Predictor \| reentraine de zero sur les 524 voix, 15 000 steps (loss 1.3 -> 0.059) \|
	\| TTL \| finetune, batch 7 x accumulation 8 (batch effectif 56) \|
	\| LR \| 2.5e-4, reduit a 1.25e-4 apres plateau de loss + baisse de `sim_ecapa` \|
	\| Checkpoints \| tous les 1 000 steps, exportes en ONNX et evalues automatiquement \|

	Pourquoi le DP est reentraine : c'est lui qui portait le begaiement du run 4-voix
	(il est conditionne par le locuteur). Sur 524 voix, il apprend un rythme robuste.

	## Utilisation (ONNX, CPU)

	```python
	from huggingface_hub import snapshot_download
	d = snapshot_download("mimba/bluetts-nnh", allow_patterns=["eval/step_9000/onnx/*"])
	# -> text_encoder / vector_estimator / vocoder / duration_predictor
	# + codec_encoder / style_encoder / duration_style_encoder (clonage zero-shot)
	```

	Parametres de production (worker Mimba) : `total_step=8..16`, `speed=0.95`,
	`cfg_scale=4.0`, `pace_blend=0.30`. Normaliser la sortie avant ecriture
	(RMS 0.08 / peak 0.95) : le vocodeur sort a un pic ~1.5 et s'ecrete sinon.

	Le vocabulaire BlueTTS ne contient pas les marqueurs de prenasalisation ngiemboons
	`m` en exposant et `n` en exposant : les remplacer par `m` / `n` avant encodage.

	## Limites

	- Fidelite de clonage inferieure a la base : `sim_ecapa` passe de ~0.35 (step 0)
	a ~0.32 apres finetune. La degradation a ete stoppee (LR reduit), pas annulee.
	- Variable selon la voix : certaines references se clonent nettement moins bien
	(voir l'ecart min/max du tableau) — tester avec ses propres references.
	- Monolingue : entraine pour le ngiemboon uniquement.
	- Donnees synthetiques : le corpus multi-voix est genere (OmniVoice + OpenVoice),
	seul l'accent provient de locuteurs natifs reels.

	## Licence

	`cc-by-nc-sa-4.0` — usage non commercial. Verifier independamment les licences de
	BlueTTS, OmniVoice et OpenVoice avant tout usage commercial.

	##### Contact : [@Mimba](baounabaouna@gmail.com)