arsalan-anwari/kana-sounds

Kana Sounds

104 short spoken clips, one for every hiragana and katakana character used by Kana Trainer: the 46 seion, 20 dakuon, 5 handakuon and 33 yoon. Every clip is trimmed, pitch corrected and loudness normalised to a single voice level, so a set can be played back to back without any jump in volume.

Dataset structure

audio/
  seion/       46 clips   a.mp3, i.mp3, ka.mp3, ... n.mp3
  dakuon/      20 clips   ga.mp3, za.mp3, ji.mp3, ... bo.mp3
  handakuon/    5 clips   pa.mp3, pi.mp3, pu.mp3, pe.mp3, po.mp3
  yoon/        33 clips   kya.mp3, sya.mp3, cya.mp3, ... pyo.mp3
favicon.png    the Kana Trainer icon

A clip is addressed by its group and romaji, audio/<group>/<romaji>.mp3. Hiragana and katakana share a clip, because the two scripts are read out the same way: audio/seion/a.mp3 is the sound of both あ and ア.

Audio format

Format	MP3 (LAME), 128 kbps
Channels	mono
Sample rate	44.1 kHz
Clips	104
Duration	0.39 s to 1.54 s, 0.77 s median, 81 s in total
Loudness	-20 dBFS RMS, peaks held under -1 dBFS
Pitch	shifted towards 330 Hz where the source allowed it
Total size	1.6 MB

Processing

The source recordings drift from 149 Hz to 367 Hz and vary about 15 dB in level. Each one is put through tools/normalize-audio.py:

1. Trim leading and trailing silence at 35 dB below the loudest frame, keep 30 ms of head and 60 ms of tail padding, and apply an 8 ms fade at both ends.
2. Pitch estimate the fundamental by autocorrelation and shift towards 330 Hz with sox, clamped to 6 semitones so a badly off source is not mangled.
3. Level scale to -20 dBFS RMS, then pull the whole clip down if any peak passes -1 dBFS.
4. Encode back to mono 128 kbps MP3 with ffmpeg.

Naming

File names follow the romaji spelling Kana Trainer shows, with four exceptions where the clip keeps the kunrei style name of the source recording:

Kana	Shown as	File
ぢ / ヂ	dji	audio/dakuon/di.mp3
づ / ヅ	dzu	audio/dakuon/du.mp3
しゃ しゅ しょ	sha shu sho	audio/yoon/sya.mp3, syu, syo
ちゃ ちゅ ちょ	cha chu cho	audio/yoon/cya.mp3, cyu, cyo
じゃ じゅ じょ	ja ju jo	audio/yoon/zya.mp3, zyu, zyo

Usage

from datasets import load_dataset

ds = load_dataset("audiofolder", data_files="hf://datasets/arsalan-anwari/kana-sounds/audio/**")

Or grab a single clip:

from huggingface_hub import hf_hub_download

path = hf_hub_download("arsalan-anwari/kana-sounds", "audio/seion/a.mp3", repo_type="dataset")

Character map

All 104 characters and the clip each one plays

Seion (basic sounds)

Row	Hiragana	Katakana	Romaji	File
A	あ	ア	a	audio/seion/a.mp3
A	い	イ	i	audio/seion/i.mp3
A	う	ウ	u	audio/seion/u.mp3
A	え	エ	e	audio/seion/e.mp3
A	お	オ	o	audio/seion/o.mp3
KA	か	カ	ka	audio/seion/ka.mp3
KA	き	キ	ki	audio/seion/ki.mp3
KA	く	ク	ku	audio/seion/ku.mp3
KA	け	ケ	ke	audio/seion/ke.mp3
KA	こ	コ	ko	audio/seion/ko.mp3
SA	さ	サ	sa	audio/seion/sa.mp3
SA	し	シ	shi	audio/seion/shi.mp3
SA	す	ス	su	audio/seion/su.mp3
SA	せ	セ	se	audio/seion/se.mp3
SA	そ	ソ	so	audio/seion/so.mp3
TA	た	タ	ta	audio/seion/ta.mp3
TA	ち	チ	chi	audio/seion/chi.mp3
TA	つ	ツ	tsu	audio/seion/tsu.mp3
TA	て	テ	te	audio/seion/te.mp3
TA	と	ト	to	audio/seion/to.mp3
NA	な	ナ	na	audio/seion/na.mp3
NA	に	ニ	ni	audio/seion/ni.mp3
NA	ぬ	ヌ	nu	audio/seion/nu.mp3
NA	ね	ネ	ne	audio/seion/ne.mp3
NA	の	ノ	no	audio/seion/no.mp3
HA	は	ハ	ha	audio/seion/ha.mp3
HA	ひ	ヒ	hi	audio/seion/hi.mp3
HA	ふ	フ	fu	audio/seion/fu.mp3
HA	へ	ヘ	he	audio/seion/he.mp3
HA	ほ	ホ	ho	audio/seion/ho.mp3
MA	ま	マ	ma	audio/seion/ma.mp3
MA	み	ミ	mi	audio/seion/mi.mp3
MA	む	ム	mu	audio/seion/mu.mp3
MA	め	メ	me	audio/seion/me.mp3
MA	も	モ	mo	audio/seion/mo.mp3
YA	や	ヤ	ya	audio/seion/ya.mp3
YA	ゆ	ユ	yu	audio/seion/yu.mp3
YA	よ	ヨ	yo	audio/seion/yo.mp3
RA	ら	ラ	ra	audio/seion/ra.mp3
RA	り	リ	ri	audio/seion/ri.mp3
RA	る	ル	ru	audio/seion/ru.mp3
RA	れ	レ	re	audio/seion/re.mp3
RA	ろ	ロ	ro	audio/seion/ro.mp3
WA	わ	ワ	wa	audio/seion/wa.mp3
WA	を	ヲ	wo	audio/seion/wo.mp3
N	ん	ン	n	audio/seion/n.mp3

Dakuon (voiced)

Row	Hiragana	Katakana	Romaji	File
GA	が	ガ	ga	audio/dakuon/ga.mp3
GA	ぎ	ギ	gi	audio/dakuon/gi.mp3
GA	ぐ	グ	gu	audio/dakuon/gu.mp3
GA	げ	ゲ	ge	audio/dakuon/ge.mp3
GA	ご	ゴ	go	audio/dakuon/go.mp3
ZA	ざ	ザ	za	audio/dakuon/za.mp3
ZA	じ	ジ	ji	audio/dakuon/ji.mp3
ZA	ず	ズ	zu	audio/dakuon/zu.mp3
ZA	ぜ	ゼ	ze	audio/dakuon/ze.mp3
ZA	ぞ	ゾ	zo	audio/dakuon/zo.mp3
DA	だ	ダ	da	audio/dakuon/da.mp3
DA	ぢ	ヂ	dji	audio/dakuon/di.mp3
DA	づ	ヅ	dzu	audio/dakuon/du.mp3
DA	で	デ	de	audio/dakuon/de.mp3
DA	ど	ド	do	audio/dakuon/do.mp3
BA	ば	バ	ba	audio/dakuon/ba.mp3
BA	び	ビ	bi	audio/dakuon/bi.mp3
BA	ぶ	ブ	bu	audio/dakuon/bu.mp3
BA	べ	ベ	be	audio/dakuon/be.mp3
BA	ぼ	ボ	bo	audio/dakuon/bo.mp3

Handakuon (p sounds)

Row	Hiragana	Katakana	Romaji	File
PA	ぱ	パ	pa	audio/handakuon/pa.mp3
PA	ぴ	ピ	pi	audio/handakuon/pi.mp3
PA	ぷ	プ	pu	audio/handakuon/pu.mp3
PA	ぺ	ペ	pe	audio/handakuon/pe.mp3
PA	ぽ	ポ	po	audio/handakuon/po.mp3

Yoon (contracted)

Row	Hiragana	Katakana	Romaji	File
KYA	きゃ	キャ	kya	audio/yoon/kya.mp3
KYA	きゅ	キュ	kyu	audio/yoon/kyu.mp3
KYA	きょ	キョ	kyo	audio/yoon/kyo.mp3
SHA	しゃ	シャ	sha	audio/yoon/sya.mp3
SHA	しゅ	シュ	shu	audio/yoon/syu.mp3
SHA	しょ	ショ	sho	audio/yoon/syo.mp3
CHA	ちゃ	チャ	cha	audio/yoon/cya.mp3
CHA	ちゅ	チュ	chu	audio/yoon/cyu.mp3
CHA	ちょ	チョ	cho	audio/yoon/cyo.mp3
NYA	にゃ	ニャ	nya	audio/yoon/nya.mp3
NYA	にゅ	ニュ	nyu	audio/yoon/nyu.mp3
NYA	にょ	ニョ	nyo	audio/yoon/nyo.mp3
HYA	ひゃ	ヒャ	hya	audio/yoon/hya.mp3
HYA	ひゅ	ヒュ	hyu	audio/yoon/hyu.mp3
HYA	ひょ	ヒョ	hyo	audio/yoon/hyo.mp3
MYA	みゃ	ミャ	mya	audio/yoon/mya.mp3
MYA	みゅ	ミュ	myu	audio/yoon/myu.mp3
MYA	みょ	ミョ	myo	audio/yoon/myo.mp3
RYA	りゃ	リャ	rya	audio/yoon/rya.mp3
RYA	りゅ	リュ	ryu	audio/yoon/ryu.mp3
RYA	りょ	リョ	ryo	audio/yoon/ryo.mp3
GYA	ぎゃ	ギャ	gya	audio/yoon/gya.mp3
GYA	ぎゅ	ギュ	gyu	audio/yoon/gyu.mp3
GYA	ぎょ	ギョ	gyo	audio/yoon/gyo.mp3
JA	じゃ	ジャ	ja	audio/yoon/zya.mp3
JA	じゅ	ジュ	ju	audio/yoon/zyu.mp3
JA	じょ	ジョ	jo	audio/yoon/zyo.mp3
BYA	びゃ	ビャ	bya	audio/yoon/bya.mp3
BYA	びゅ	ビュ	byu	audio/yoon/byu.mp3
BYA	びょ	ビョ	byo	audio/yoon/byo.mp3
PYA	ぴゃ	ピャ	pya	audio/yoon/pya.mp3
PYA	ぴゅ	ピュ	pyu	audio/yoon/pyu.mp3
PYA	ぴょ	ピョ	pyo	audio/yoon/pyo.mp3

Licence and credits

The recordings come from Learn Japanese Adventure and are used under CC BY 4.0; the normalised clips in this dataset carry the same licence. Please keep the attribution when you reuse them.

The Kana Trainer application itself is Apache-2.0 and lives at github.com/arsalan-anwari/kana-trainer.

Citation

@misc{anwari_kana_sounds,
  title  = {Kana Sounds},
  author = {Anwari, Arsalan},
  url    = {https://huggingface.co/datasets/arsalan-anwari/kana-sounds},
  note   = {Recordings from Learn Japanese Adventure, CC BY 4.0}
}