Multispeaker Community Vocoder model for DiffSinger

Trained with ~95 hours of varied singing data.

The goal of our vocoder model is to provide more quality possibilities that can be brought out of DiffSinger acoustics. The vocoder can be used with any voice.
We would like to send huge thanks to the Western Diffsinger community for providing datasets for training! Without them, this model wouldn’t be possible.

Code used to train the vocoder is avaible on HiFiPLN Github.
Pretrained checkpoint for finetuning is available at the bottom of this page.

• Model Developed & Trained by Scarfmonster
• Data coordination by PixPrucer

How To Use

1. Download your vocoder of choice
2. Drag and drop the downloaded .oudep file onto the OpenUtau window.
3. Change the dsconfig.yaml configuration file of your chosen voice and set vocoder: to a proper value. For your convenienve the setting is listed in the table. Save the file, restart OpenUtau if you had it opened.

Download

Name	Download	Version	dsconfig.yaml	Notes
hifipln	89,1 MB | --	1.1 (2024-02-17)	vocoder: hifipln_1.1	sample_rate: 44100, n_mels: 128, hop_length: 512
ddsppln	23.2 MB | --	1.0 (2024-01-29)	vocoder: ddsppln_1.0	A DDSP-like vocoder. Not realistic, somewhat robotic, but some people like the sound. sample_rate: 44100, n_mels: 128, hop_length: 512
hifipln	89,1 MB | --	1.0 (2024-02-03)	vocoder: hifipln_1.0	sample_rate: 44100, n_mels: 128, hop_length: 512

Singing databases used for training the model

Name	Length	Languages	Contributor
AdoVoc Pro	00:05:28	Caló, Spanish	AdoVoc Pro
A.I.chi	01:47:52	English, Japanese	Peeslubn
Aida	04:00:48	EN, JA, DE, FR	Violin
Albert	01:17:14	Polish	SzTJ
Aleks	00:06:10	Polish	SzTJ
Ameko Kero	02:54:31	English, Japanese	HoodyPisDed
Ariel	01:09:01	Japanese	ariika
Brent	00:05:15	Spanish	Beatrix
Cantoria Dataset	02:25:08	Spanish	Cantoria Dataset
Codie	01:00:37	Japanese	code41den
Deshi	01:47:45	Japanese, Tagalog	UtaUtaUtau
Esmuc Choir Dataset	00:21:31	German	Esmuc Choir Dataset
Evelyn	01:05:32	English	Violin
Filip	01:17:49	Polish	Rainygardens
Geppei	00:30:15	Japanese, Polish, Ukrainian	vahntanabe
Hania	00:05:32	Polish	SzTJ
Hisaki	02:42:57	Japanese	ryutsu
Inka	00:39:18	English, Japanese	postTEENIDOL
Jalo	00:54:53	Polish	SzTJ
Karasu	00:49:55	Japanese	rev
Kazuo	00:33:40	Japanese	Felipe Souza
Kiiro	01:44:54	English, Japanese	Ryouichi
Konryuu	01:10:55	Japanese	PixPrucer
Kurenai	00:55:26	Japanese	liure
Leif	01:28:40	English, Japanese	Tigermeat
Lem	00:14:22	Polish	Wik
Liee	00:25:01	JA, EN, Latin	julieraptor
Makam Acapella	00:38:53	Turkish	Makam Acapella
Makku	02:06:58	JA, EN, ES, IT	Gianloop
Mat	00:35:44	Polish	hq_png
Matsuki Max	01:25:32	Japanese	Haraoo
Mava	01:46:33	English, Japanese	Enzo
Mora	01:49:03	English, Japanese	funhouse
Namine Criss	00:31:02	Spanish, Japanese	CrissZ3R0VZ
Nanabot	00:29:23	English	postTEENIDOL
Naoky	03:31:55	EN, JA, KO, ZH	xuu
——	03:55:02	——	Anonymous Contributor
Paulina	00:29:34	Polish	SzTJ
Peiton	02:31:09	English	NebulaMeadow
PIX	04:10:54	Polish, Japanese	PixPrucer
Otozora Rinly	02:49:43	Japanese	UniverStars
Ron	02:28:10	EN, JA, PL, KO, ZH	Galanist
Rose	00:42:39	Polish, Japanese	Kisa
Ryszard	02:24:16	Polish, Japanese	Scarfmonster
——	01:50:00	——	Anonymous Contributor
Singing Database	02:46:46	Chinese, Italian	Singing Database
Ace	02:50:26	English, Japanese	SpoopyAce
Stefan	02:49:07	Polish, Latin	SzTJ
Suzu	01:42:03	Japanese	ariika
Taylor	01:09:24	English	postTEENIDOL
Teo Vampa	01:56:33	Japanese	Delphic
Tetsu	01:13:46	Japanese	ariika
Tiger	03:31:27	EN, ES, JP, KO, ZH, PT, FR	Tigermeat
Tomo	00:57:00	Spanish, Japanese	Tomo
Vocadito	00:13:37	EN, FR, HAW, ES, TL, Valencian	Vocadito
VocalSet	08:46:18	Vocalise	VocalSet
Wanda	01:11:03	Polish	Vieri
Wioletta	00:32:56	Polish	SzTJ
Zethiel Yu	02:19:19	English	xiel exalt
Zethiel Zero	00:32:07	English, Japanese	xiel exalt
Total length:	98:28:51
Used length:	82:06:15

Pitch distribution

Dataset

After augmentation

Checkpoints for finetuning

Name	Download	Version	Notes
hifipln	378 MB | --	1.0 (2024-02-03)	sample_rate: 44100, n_mels: 128, hop_length: 512