Dataset Viewer
Auto-converted to Parquet Duplicate
id
string
text_id
string
voice_id
string
text
string
split
string
category
string
origin
string
source_id
string
source_license
string
source_url
string
audio
dict
duration_seconds
float64
sample_rate
int32
asr_transcript
string
normalized_cer
float64
attempt
int32
generation_seed
int64
generation_batch
int32
teacher_model
string
reference_audio
string
review_reasons
list
cv-00003_F2
cv-00003
F2
"「ある」とは、ロッツェがいったように、「関係に立つ」ということであ(...TRUNCATED)
train
general
common_voice
sentence-collector.txt line 3
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAAIcABgaBdwA8AAC9YAW8ymQJlTaJeqvIwsBENOshAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
8.08
24,000
"あるとは、ロッツが言ったように、関係に立つということであり、関係に(...TRUNCATED)
0.0196
1
20,363,217
2,103
Qwen/Qwen3-TTS-12Hz-1.7B-Base
F2.wav
[]
cv-00003_M2
cv-00003
M2
"「ある」とは、ロッツェがいったように、「関係に立つ」ということであ(...TRUNCATED)
train
general
common_voice
sentence-collector.txt line 3
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAAMKABUvBdwA8AACdgBcwA0/8FksSV/2ryxQTcBRhAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
6.72
24,000
"あるとは、ロッツェが言ったように、関係に立つということであり、関係(...TRUNCATED)
0
1
20,363,217
2,103
Qwen/Qwen3-TTS-12Hz-1.7B-Base
M2.wav
[]
cv-00005_F2
cv-00005
F2
「はい、出来ます。」と、少年は答えました。
train
general
common_voice
sentence-collector.txt line 5
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAAEwABqfBdwA8AABUYCtNo9vZEM6ydWoqNfxrlYDhAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
3.6
24,000
はい、できます。と少年は答えました。
0
1
20,362,488
1,374
Qwen/Qwen3-TTS-12Hz-1.7B-Base
F2.wav
[]
cv-00005_M2
cv-00005
M2
「はい、出来ます。」と、少年は答えました。
train
general
common_voice
sentence-collector.txt line 5
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAAAvABnCBdwA8AABYICAHy38B+8dQY0UVuNDpyhYhAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
3.76
24,000
はい、できます。と少年は答えました。
0
1
20,362,488
1,374
Qwen/Qwen3-TTS-12Hz-1.7B-Base
M2.wav
[]
cv-00010_F2
cv-00010
F2
"「眠っちゃだめだ。」と、少年は思いました。「眠ったら、午前中にお説(...TRUNCATED)
train
general
common_voice
sentence-collector.txt line 10
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAAQRABkmBdwA8AAC/QB/W+cD4fsVCw2SKo+tTw+ZhAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
8.16
24,000
"眠っちゃダメだ、と、少年は思いました。眠ったら、午前中にお説教は読(...TRUNCATED)
0
1
20,363,182
2,068
Qwen/Qwen3-TTS-12Hz-1.7B-Base
F2.wav
[]
cv-00010_M2
cv-00010
M2
"「眠っちゃだめだ。」と、少年は思いました。「眠ったら、午前中にお説(...TRUNCATED)
train
general
common_voice
sentence-collector.txt line 10
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAAP9ABg3BdwA8AAC/QDgWlEvzehBnZ6VnLdL/bxwhAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
8.16
24,000
"眠っちゃダメだ。と、少年は思いました。眠ったら、午前中におセッキュ(...TRUNCATED)
0.0408
1
20,363,182
2,068
Qwen/Qwen3-TTS-12Hz-1.7B-Base
M2.wav
[]
cv-00016_F2
cv-00016
F2
ああ、わかってるさ。
train
general
common_voice
sentence-collector.txt line 16
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAAO3ABk/BdwA8AAA2YDBZpe57z4UGXsebAKhXgUvhAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
2.32
24,000
ああ、分かってるさ
0
1
20,361,230
116
Qwen/Qwen3-TTS-12Hz-1.7B-Base
F2.wav
[]
cv-00016_M2
cv-00016
M2
ああ、わかってるさ。
train
general
common_voice
sentence-collector.txt line 16
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAABAABUYBdwA8AAA4QCmJTO7O3NBopqBU0+KxBrshAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
2.4
24,000
ああ、わかってるさ
0
1
20,361,230
116
Qwen/Qwen3-TTS-12Hz-1.7B-Base
M2.wav
[]
cv-00023_M2
cv-00023
M2
あいつは言い放った。
train
general
common_voice
sentence-collector.txt line 23
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAAHAABbyBdwA8AAApQDot9Qpv2hc2rHHdBJakcHMhAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
1.76
24,000
あいつは言い放った
0
1
20,361,230
116
Qwen/Qwen3-TTS-12Hz-1.7B-Base
M2.wav
[]
cv-00034_F2
cv-00034
F2
あしたからやる。
train
general
common_voice
sentence-collector.txt line 34
CC0-1.0
https://github.com/common-voice/common-voice/blob/main/server/data/ja/sentence-collector.txt
{"bytes":"ZkxhQwAAACIQABAAAAIaABjyBdwA8AAApQCYs9ERPFk3AO+lPlTI93mZhAAAKCAAAAByZWZlcmVuY2UgbGliRkxBQy(...TRUNCATED)
1.76
24,000
明日からやる
0
1
20,361,151
37
Qwen/Qwen3-TTS-12Hz-1.7B-Base
F2.wav
[]
End of preview. Expand in Data Studio

tiny-voca-ja dataset

日本語要約: 英語 TTS モデル tiny-voca の学習用コーパスを日本語で作り直した 合成音声データセットです。文章 8,500 文を Qwen3-TTS で 2 声(F2/M2)に読み上げ、 Whisper の書き起こしで選別した 16,321 クリップ(16.6 時間)を収録しています。 実在の話者の音声は含まれません。文章のライセンスは行ごとに異なります (Tatoeba は CC BY 2.0 FR で出典表示が必要、Common Voice は CC0)。 詳細は License の節を見てください。

A synthetic Japanese speech corpus with two voices, F2 (female) and M2 (male), made to train the tiny-voca-ja text-to-speech model. It is the Japanese counterpart of the tiny-voca dataset: the same production method, with Japanese text sources. Every sentence was meant to be read by both voices; sentences where one reading failed the screening appear with one voice only.

Clips 16,321
Duration 16.562 hours
Voices F2, M2
Audio 24 kHz mono, 16-bit FLAC, stored inside the Parquet files
Language Japanese
License Per row — see the source_license column and the License section below. There is no single license for the whole dataset.

Usage

from datasets import load_dataset

ds = load_dataset("gachimoto/tiny-voca-ja")

row = ds["train"][0]
print(row["voice_id"], row["text"], row["source_license"])
audio = row["audio"]["array"]

Contents

Splits are made by sentence: both readings of a sentence are always in the same split.

Split Clips Hours
train 15,366 15.569
validation 479 0.491
test 476 0.502

How it was made

  1. The two voices were created from text descriptions with Qwen3-TTS 1.7B VoiceDesign. They are not recordings or imitations of real people. The reference clips are in references/, and the descriptions are in voices.json.
  2. Each of the 8,500 sentences was synthesized in both voices with Qwen3-TTS 1.7B Base, using the reference clips as the voice prompt (17,000 clips targeted).
  3. Each clip was transcribed with Whisper small (multilingual). Clips whose kana-normalized character error rate against the text exceeded 15%, or with implausible duration, speaking rate or silence, were generated a second time. Clips that failed twice were left out: 679 of 17,000. They are listed in generation/review.jsonl; the accepted clips, with full provenance, are in generation/manifest.jsonl.
  4. Accepted clips: 16,321 of 17,000 (96.0%). Mean screening CER of the accepted clips is 1.76%.

The automatic screening catches wrong or missing words. It does not measure naturalness, and the corpus has not been checked by listeners clip by clip.

Text sources

The origin, source_id, source_license and source_url columns record the source of every row.

Source Sentences in corpus Accepted clips Accepted sentences License of the text
Tatoeba (Japanese sentences) 5,900 11,420 5,778 CC BY 2.0 FR — attribution required
Mozilla Common Voice sentence collector (Japanese) 2,000 3,816 1,942 CC0
Templates written for this project: numbers, dates, money, names, questions 600 1,085 563 Written for this project

Columns

Column Description
id Clip ID, <text_id>_<voice_id>
text_id, voice_id Sentence and voice
text The sentence that was read
split train, validation or test
category, origin Kind of sentence and where it came from
source_id, source_license, source_url Source attribution for the text
audio 24 kHz FLAC audio
duration_seconds, sample_rate Audio length and rate
asr_transcript, normalized_cer Whisper transcript and its kana-normalized character error rate against text
attempt, generation_seed, generation_batch Generation attempt (1 or 2), random seed and batch
teacher_model, reference_audio Model and voice prompt used
review_reasons Always empty for accepted clips (the reasons live in generation/review.jsonl)

Annotations

annotations/beatrice/ holds the phoneme-level annotations used for training: phonemes, stress, and the duration of each phoneme in 10.7 ms frames, from forced alignment with prj-beatrice/japanese-hubert-base-phoneme-ctc-v2 (labels from pyopenjtalk-plus g2p, transferred at mora level).

Split Accepted Of
train 10,884 15,366
validation 332 479
test 345 476
total 11,561 (70.8%) 16,321

Clips that failed the alignment checks are listed with a reason in the *.rejected.jsonl files. Most rejections are reading mismatches (kanji readings, rendaku) between the label sequence and the audio.

Note: the English dataset's aligner (facebook/wav2vec2-lv-60-espeak-cv-ft) accepted 0 of 150 Japanese smoke-test clips, which is why this dataset uses the beatrice aligner instead.

Limitations

  • The speech is synthetic. It carries the pronunciation habits and errors of the model that generated it.
  • Selection bias: sentences the teacher model misread — mostly numbers, dates and proper nouns — failed the screening disproportionately, so those kinds of sentences are underrepresented in the accepted clips. This is visible downstream: the trained model is weakest exactly on numbers and dates.
  • Two voices, one reading style, read speech only.
  • Screening is ASR-based only; no human listening check per clip.

License

There is no single license for this dataset, because the text sources have different terms. Per row:

  • Rows with source_license CC-BY-2.0-FR contain sentences from Tatoeba, © their respective contributors, licensed under CC BY 2.0 FR. Attribution is required: credit Tatoeba and link to the sentence — each row's source_url does this.
  • Rows with source_license CC0-1.0 contain sentences from the Mozilla Common Voice sentence collector (CC0).
  • Rows with source_license project-own are template sentences written for this project.

The audio was generated with Qwen3-TTS (Apache-2.0) from designed synthetic voices; no real person's voice is included. The compilation of this dataset, the project-written template sentences and the annotations are released by the publisher under CC BY 4.0. See LICENSE.md for the full text of these terms.

This differs from the English tiny-voca dataset, which is CC BY 4.0 as a whole: its book text was already CC BY 4.0, while the Japanese text sources include CC BY 2.0 FR material that keeps its own license.

@misc{fujimoto2026tinyvocajadata,
  author = {Fujimoto, Satoshi},
  title  = {tiny-voca-ja dataset: synthetic two-voice Japanese speech},
  year   = {2026},
  url    = {https://huggingface.co/datasets/gachimoto/tiny-voca-ja}
}

Author: gachimoto (Satoshi Fujimoto).

The Tatoeba sentences are used under CC BY 2.0 FR: https://tatoeba.org — please keep the per-sentence attribution in source_id / source_url when redistributing. The Common Voice sentences are CC0, from https://github.com/common-voice/common-voice. The audio was generated with Qwen3-TTS (Apache-2.0). See LICENSE.md.

Downloads last month
20

Models trained or fine-tuned on gachimoto/tiny-voca-ja