inklight-app/paced-voices

Inklight paced voices

Voice states for Kyutai's Pocket TTS, made for Inklight, an iPhone app that reads books aloud on the device. Each file is a voice state for the english_2026-04 model: the language model's key-value cache after hearing a speaker's reference clip. Most are slower states: the clip was time-stretched before the model heard it, and the model imitates the clip's tempo, so a voice loaded from one of these files reads more slowly of its own accord, with the pauses a slow reader adds, and no time-stretching of the output. Eight voices that Kyutai does not ship also have their state at their own pace here.

Files

english_2026-04/<voice>@<pace>.safetensors, four paces for each of eighteen voices, 72 files, 7 to 12.5 MB each. The pace in the name is the speed the state was measured to produce relative to the voice at its own pace (0.9, 0.85, 0.8 and 0.75); every file was measured within about 0.02 of its name, over three sentences and three or six seeds, with the articulation rate taken from the rendered speech and pauses left out. stops.json lists, for every file, the clip stretch it was made from, the measured pace and its spread, and, where one was made, the word error rate of a check transcription.

english_2026-04/<voice>.safetensors, the state at the voice's own pace, for the eight voices named by their VCTK speaker ids (p227, p238, p263, p270, p272, p294, p304, p351), 6 to 8 MB each. 720 MB in all.

The other ten voice names are Kyutai's voice ids in Pocket TTS, whose states at their own pace Kyutai publishes. The files are in the same format as Kyutai's own voice embeddings for that model (a per-layer cache and its offset) and are read the same way.

How they were made

The reference clip of each voice, the same recording Kyutai made its voices from (_ORIGINS_OF_PREDEFINED_VOICES in kyutai-labs/pocket-tts; for the eight VCTK speakers Kyutai does not ship, the cleaned clip of sentence 23 in kyutai/tts-voices), was time-stretched with ffmpeg and encoded with the voice-cloning release of the model. How far a voice follows the stretch is the speaker's own, so for each voice the stretch was swept, each candidate state measured, and the one closest to each pace kept. The eight states at their own pace were encoded the same way from the unstretched clip.

2026-10-29: states at fractions of each narrator's natural pace

english_2026-04/261029/<voice>@<pace>.safetensors, 24 files for eight voices (anna, azelma, charles, eponine, michael, p270, paul, vera), 7 to 12.5 MB each. Here the pace in the name is a fraction of the voice's own natural pace (0.9, 0.8, 0.7 and, for charles only, 0.6), measured in words per minute of voiced time on a 256-word passage over four seeds; every file reads within 3 % of its name, all but two within 1.5 %. The clips were slowed with ffmpeg's atempo for charles, michael, paul and p270 and with Praat's PSOLA for the other four, by factors between 0.50 and 0.88. A pace was kept only where the voice stays recognisably the speaker's own (ECAPA speaker similarity of at least 0.85 against the voice at its own pace) and the check transcription stays clean; the slower paces that missed this are not published. The files in english_2026-04/ itself are unchanged.

Credits and licence

These states are adaptations of recordings published under the Creative Commons Attribution 4.0 International licence, and are released under the same licence.

• anna, azelma, charles, eponine, fantine, george, michael, paul and vera are derived from the CSTR VCTK Corpus, version 0.92, by Junichi Yamagishi, Christophe Veaux and Kirsten MacDonald, The Centre for Speech Technology Research (CSTR), University of Edinburgh, doi.org/10.7488/ds/2645 (speakers p228, p303, p254, p262, p244, p315, p360, p259 and p229), licensed under CC BY 4.0, by way of the cleaned recordings in kyutai/tts-voices.
• p227, p238, p263, p270, p272, p294, p304 and p351 are derived from the same corpus, speakers of the same numbers, by way of the same cleaned recordings.
• alba is derived from the recording "Casual" by Alba MacKenna, kyutai/tts-voices, licensed under CC BY 4.0.
• The model that computed them is Pocket TTS by Kyutai, licensed under CC BY 4.0.

Changes made: each recording was slowed, without changing its pitch, to between 0.6 and 0.9 of its original speed, and reduced to a Pocket TTS voice state; for the eight VCTK speakers above, the recording was also reduced to a voice state unslowed. Neither the speakers, CSTR, nor Kyutai endorse these files or the app. They are provided without warranty.