Dataset Viewer
Auto-converted to Parquet Duplicate
audio
audioduration (s)
0.48
1.89
text
stringclasses
7 values
engine
stringclasses
4 values
voice
stringlengths
14
100
speed
stringclasses
221 values
lang
stringclasses
1 value
rate
stringclasses
236 values
pitch
stringclasses
1 value
speaker_id
stringclasses
3 values
source_clip
stringclasses
422 values
origin
stringclasses
5 values
variety
stringclasses
1 value
seed
stringlengths
0
12
reference
stringclasses
623 values
asr_text
stringclasses
1 value
qa
stringclasses
2 values
reject_reason
stringclasses
1 value
esnatu
omnivoice
auto; seed9449790000-item0
0.9
eu
esnatu!
omnivoice
auto; seed9449790000-item1
0.9
eu
esnatu?
omnivoice
auto; seed9449790000-item2
0.9
eu
esnatu
omnivoice
auto; seed9449790000-item3
1.1
eu
esnatu?
omnivoice
auto; seed9449790000-item5
1.1
eu
esnatu
omnivoice
auto; seed9449790000-item6
0.9
eu
esnatu!
omnivoice
auto; seed9449790000-item7
0.9
eu
esnatu?
omnivoice
auto; seed9449790000-item8
0.9
eu
esnatu
omnivoice
auto; seed9449790000-item9
1.1
eu
esnatu
omnivoice
auto; seed9449790010-item2
0.9
eu
esnatu!
omnivoice
auto; seed9449790010-item3
0.9
eu
esnatu?
omnivoice
auto; seed9449790010-item4
0.9
eu
esnatu
omnivoice
auto; seed9449790010-item5
1.1
eu
esnatu?
omnivoice
auto; seed9449790010-item7
1.1
eu
esnatu!
omnivoice
auto; seed9449790010-item9
0.9
eu
esnatu?
omnivoice
auto; seed9449790020-item0
0.9
eu
esnatu
omnivoice
auto; seed9449790020-item4
0.9
eu
esnatu!
omnivoice
auto; seed9449790020-item5
0.9
eu
esnatu?
omnivoice
auto; seed9449790020-item6
0.9
eu
esnatu
omnivoice
auto; seed9449790030-item0
0.9
eu
esnatu!
omnivoice
auto; seed9449790030-item1
0.9
eu
esnatu?
omnivoice
auto; seed9449790030-item2
0.9
eu
esnatu!
omnivoice
auto; seed9449790030-item4
1.1
eu
esnatu?
omnivoice
auto; seed9449790030-item5
1.1
eu
esnatu!
omnivoice
auto; seed9449790030-item7
0.9
eu
esnatu
omnivoice
auto; seed9449790040-item2
0.9
eu
esnatu?
omnivoice
auto; seed9449790040-item4
0.9
eu
esnatu!
omnivoice
auto; seed9449790040-item6
1.1
eu
esnatu?
omnivoice
auto; seed9449790040-item7
1.1
eu
esnatu!
omnivoice
auto; seed9449790040-item9
0.9
eu
esnatu
omnivoice
auto; seed9449790050-item1
1.1
eu
esnatu?
omnivoice
auto; seed9449790050-item3
1.1
eu
esnatu
omnivoice
auto; seed9449790050-item4
0.9
eu
esnatu!
omnivoice
auto; seed9449790050-item5
0.9
eu
esnatu?
omnivoice
auto; seed9449790050-item6
0.9
eu
esnatu!
omnivoice
auto; seed9449790050-item8
1.1
eu
esnatu?
omnivoice
auto; seed9449790050-item9
1.1
eu
esnatu
omnivoice
auto; seed9449790060-item0
0.9
eu
esnatu!
omnivoice
auto; seed9449790060-item1
0.9
eu
esnatu?
omnivoice
auto; seed9449790060-item2
0.9
eu
esnatu
omnivoice
auto; seed9449790060-item3
1.1
eu
esnatu
omnivoice
auto; seed9449790060-item6
0.9
eu
esnatu!
omnivoice
auto; seed9449790060-item7
0.9
eu
esnatu?
omnivoice
auto; seed9449790060-item8
0.9
eu
esnatu
omnivoice
auto; seed9449790060-item9
1.1
eu
esnatu?
omnivoice
auto; seed9449790070-item1
1.1
eu
esnatu
omnivoice
auto; seed9449790070-item2
0.9
eu
esnatu!
omnivoice
auto; seed9449790070-item3
0.9
eu
esnatu?
omnivoice
auto; seed9449790070-item4
0.9
eu
esnatu
omnivoice
auto; seed9449790070-item5
1.1
eu
esnatu!
omnivoice
auto; seed9449790070-item9
0.9
eu
esnatu?
omnivoice
auto; seed9449790080-item0
0.9
eu
esnatu?
omnivoice
auto; seed9449790080-item3
1.1
eu
esnatu!
omnivoice
auto; seed9449790080-item5
0.9
eu
esnatu?
omnivoice
auto; seed9449790080-item6
0.9
eu
esnatu
omnivoice
auto; seed9449790080-item7
1.1
eu
esnatu?
omnivoice
auto; seed9449790080-item9
1.1
eu
esnatu
omnivoice
auto; seed9449790090-item0
0.9
eu
esnatu!
omnivoice
auto; seed9449790090-item1
0.9
eu
esnatu?
omnivoice
auto; seed9449790090-item2
0.9
eu
esnatu
omnivoice
auto; seed9449790090-item3
1.1
eu
esnatu
omnivoice
auto; seed9449790090-item6
0.9
eu
esnatu!
omnivoice
auto; seed9449790090-item7
0.9
eu
esnatu?
omnivoice
auto; seed9449790090-item8
0.9
eu
esnatu
omnivoice
auto; seed9449790090-item9
1.1
eu
esnatu!
omnivoice
auto; seed9449790100-item0
1.1
eu
esnatu?
omnivoice
auto; seed9449790100-item1
1.1
eu
esnatu
omnivoice
auto; seed9449790100-item2
0.9
eu
esnatu!
omnivoice
auto; seed9449790100-item3
0.9
eu
esnatu?
omnivoice
auto; seed9449790100-item4
0.9
eu
esnatu
omnivoice
auto; seed9449790100-item5
1.1
eu
esnatu
omnivoice
auto; seed9449790100-item8
0.9
eu
esnatu?
omnivoice
auto; seed9449790110-item0
0.9
eu
esnatu
omnivoice
auto; seed9449790110-item1
1.1
eu
esnatu!
omnivoice
auto; seed9449790110-item2
1.1
eu
esnatu?
omnivoice
auto; seed9449790110-item3
1.1
eu
esnatu?
omnivoice
auto; seed9449790110-item6
0.9
eu
esnatu
omnivoice
auto; seed9449790110-item7
1.1
eu
esnatu!
omnivoice
auto; seed9449790110-item8
1.1
eu
esnatu?
omnivoice
auto; seed9449790110-item9
1.1
eu
esnatu!
omnivoice
auto; seed9449790120-item1
0.9
eu
esnatu?
omnivoice
auto; seed9449790120-item2
0.9
eu
esnatu?
omnivoice
auto; seed9449790120-item5
1.1
eu
esnatu
omnivoice
auto; seed9449790120-item6
0.9
eu
esnatu!
omnivoice
auto; seed9449790120-item7
0.9
eu
esnatu?
omnivoice
auto; seed9449790120-item8
0.9
eu
esnatu
omnivoice
auto; seed9449790120-item9
1.1
eu
esnatu!
omnivoice
auto; seed9449790130-item3
0.9
eu
esnatu?
omnivoice
auto; seed9449790130-item4
0.9
eu
esnatu
omnivoice
auto; seed9449790130-item8
0.9
eu
esnatu!
omnivoice
auto; seed9449790130-item9
0.9
eu
esnatu?
omnivoice
auto; seed9449790140-item0
0.9
eu
esnatu
omnivoice
auto; seed9449790140-item1
1.1
eu
esnatu!
omnivoice
auto; seed9449790140-item2
1.1
eu
esnatu
omnivoice
auto; seed9449790140-item4
0.9
eu
esnatu!
omnivoice
auto; seed9449790140-item5
0.9
eu
esnatu!
omnivoice
auto; seed9449790140-item8
1.1
eu
esnatu!
omnivoice
auto; seed9449790150-item1
0.9
eu
esnatu?
omnivoice
auto; seed9449790150-item2
0.9
eu
esnatu!
omnivoice
auto; seed9449790150-item4
1.1
eu
End of preview. Expand in Data Studio

synthetic-wakeword-esnatu

Synthetic wake-word audio for training and benchmarking OVOS wake-word plugins, covering the phrase "esnatu".

Every clip is machine-generated: by a text-to-speech model with no reference speaker, or, in chatterbox-vc/, by converting such a clip to the voice of a public Common Voice reference recording, so the clips reproduce no private voice. The dataset is published under CC BY 4.0, free to use, redistribute and build on, including for model training.

Produced with support from the NGI0 Commons Fund.

Layout

Training clips are in train/ (3,428 clips), inside a folder named after how it was made; the folders are described below. Each split has a metadata.jsonl giving each clip's text, engine, voice, speed, language, seed, source clip, reference and quality-check status. The test split (test/, 291 clips) holds OmniVoice clips whose seeds are used by no training clip; it is the held-out test set of the wake-word models trained on train/. unchecked/ (512 clips) holds clips that no quality check has covered. rejected/ (959 clips) holds clips that failed the checks, with the reason in reject_reason.

omnivoice/

128 clips (of 200 generated) made with OmniVoice (k2-fsa/OmniVoice, PyTorch) with no reference speaker: every clip has a newly drawn speaker from the model's automatic voice. Generation is seeded per batch of 10 clips; voice in train/metadata.jsonl names the style and the seed. Speeds 0.9 and 1.1; text forms "esnatu", "esnatu!", "esnatu?". A clip was kept when its MMS forced-alignment score (torchaudio MMS_FA with uroman) reached the phrase's threshold (-0.636, the median of the thresholds of the other phrases, since the phrase has no edge-tts reference clips) and the checks on duration (0.4 to 5 s), level (RMS at least -45 dB) and speech activity (at least 25% active frames) passed. Each kept clip is trimmed to the aligned phrase plus 100 ms. 16 kHz mono 16-bit PCM WAV.

omnivoice-single/ (train and test)

train/omnivoice-single/ holds 1,742 clips and test/omnivoice-single/ 291 clips, made with OmniVoice (k2-fsa/OmniVoice, fp16) with no reference speaker, one clip per call, each call seeded with its own seed so that every clip has a newly drawn speaker. voice in metadata.jsonl names each clip's seed and, where one was given, its style instruction (gender, pitch, age or an English accent). Speeds 0.85 to 1.15; text forms of the phrase with and without "!" and "?". The test split holds only OmniVoice clips whose seeds no training clip uses.

Basque (eu; phrases "esnatu"): 1,742 training clips of 2,549 generated, seeds 880,009,000,000 to 880,009,002,699; 291 test clips of 443 generated, seeds 990,009,000,000 to 990,009,000,469. A clip was kept when the checks on duration (0.3 to 4 s), level (RMS at least -45 dB) and speech activity (at least 25% active frames) passed and its speech duration, scaled by its speaking rate, was within four robust standard deviations of the median (no speech recogniser covers this language).

Each kept clip is trimmed to its speech, padded with 25 to 120 ms of silence before and 40 to 350 ms after, and scaled to a level drawn from the level distribution of other synthetic wake-word test sets. 16 kHz mono 16-bit PCM WAV.

phoonnx/

320 training clips made with phoonnx text-to-speech, run through the voice models named in voice: hitz/eu-antton, hitz/eu-antton-happy, hitz/eu-antton-neutral, hitz/eu-antton-sad, hitz/eu-maider, hitz/eu-maider-happy, hitz/eu-maider-neutral, hitz/eu-maider-sad. Each clip is one voice, one text form of the phrase and one random draw of the synthesis noise scales; rate in metadata.jsonl holds the phoonnx length scale (larger is slower) and speaker_id the speaker of multi-speaker voices. Each clip is trimmed to its speech, padded and level-scaled as the omnivoice-single/ clips are and was kept when it was 0.3 to 4 s long, at least -45 dB RMS and at least 25% active frames; no recogniser checked these clips. 16 kHz mono 16-bit PCM WAV.

unchecked/

Clips that no quality check has covered. Every row of unchecked/metadata.jsonl has qa set to unchecked. These clips are not part of train/ or test/. unchecked/omnivoice-single/test-top-up/ holds 512 clips: further OmniVoice clips generated with the same recipe and seed scheme as the test split, whose seeds no training clip uses; they are raw (not trimmed and not level-scaled). The OmniVoice clips are 16 kHz mono 16-bit PCM WAV; the clips of the other generators keep the generator's native format.

rejected/

959 OmniVoice clips that failed the checks of omnivoice-single/, in rejected/omnivoice-single/train/ and rejected/omnivoice-single/test/ (the split their seeds belong to). They are the raw generator output, not trimmed and not level-scaled. reject_reason in rejected/metadata.jsonl names the failed checks (fail_asr transcript did not match the phrase, fail_dur outside 0.3 to 4 s, fail_quiet below -45 dB RMS, fail_active fewer than 25% active frames, fail_empty no speech, fail_durz speech duration an outlier) and, where a recogniser ran, its transcript.

chatterbox-vc/

938 training clips made by voice conversion with the Chatterbox model: a synthetic clip of the phoonnx/ or mms-tts/ generators converted to the voice of a reference recording. source_clip in metadata.jsonl names the converted clip, voice its original voice and reference the reference recording, a public Common Voice clip; the converted voice also carries some of the reference recording's microphone and room. Clips were trimmed, level-scaled and kept under the same duration, level and speech-activity checks as phoonnx/; no recogniser checked them. 16 kHz mono 16-bit PCM WAV.

mms-tts/

300 training clips made with Meta's MMS-TTS Basque model (facebook/mms-tts-eus), one clip per seed (seed in metadata.jsonl) with a random speaking rate (speed). Each clip is trimmed to its speech, padded and level-scaled as the omnivoice-single/ clips are and was kept when it was 0.3 to 4 s long, at least -45 dB RMS and at least 25% active frames. 16 kHz mono 16-bit PCM WAV.

Downloads last month
136

Collection including TigreGotico/synthetic-wakeword-esnatu