Say you want a phone, with no signal, to read a sentence out loud in Nepali. We forked Kyutai's Pocket TTS — a small, CPU-friendly speech model that had never heard a word of Nepali — compared it zero-shot against our own trained Nepali voice-cloning system, and it won. What we built on top of that win is a teacher, a student, and a storage problem we only partly fixed.
Our best Nepali voice model exists as one copy, on one machine.
Audio8 is our existing Nepali voice-cloning system — a 169.8M-parameter generator plus a 130M codec decoder, built the slow way on roughly 92 hours of audio across 1,300 speakers. On a 24-speaker development probe it scored an ECAPA truth-similarity cosine of 0.400 and a 23.8% character error rate, running at 0.665× real time on a four-thread Apple M3.
Same 24 voices, same texts, same scorers: Pocket's Hindi checkpoint, which has never heard a word of Nepali, against Audio8. It won — 0.487 ECAPA truth similarity against 0.400, 14.2% character error rate against 23.8%, and six times faster, 0.099× real time against 0.665×. It took 18 of 24 voices on similarity and 19 of 24 on accuracy, with the kind of statistics (paired Wilcoxon p=.0096 and p=1.3×10⁻⁴) that make you re-check your spreadsheet. The win has a catch built into it: Hindi training leaks into Nepali as mispronunciation — Hindi-shaped sounds for Nepali-shaped words. And this is a 24-speaker development probe, not a product benchmark; it supports "pick Pocket as the base," not "Pocket beats Audio8."
We trained a from-scratch bilingual Nepali/English teacher, averaged two training snapshots together — "souping" — into a 340M-parameter model: 12.3% Nepali CER, 3.5% English Seed-TTS WER. (A separate, independent harness later scored the same checkpoint under harder conditions and got far worse, inconsistent numbers, up to 95.8% CER on one test set — its own notes call that a "failed/limited" run, not a deployment pass.) We then blended the soup 30/70 with another checkpoint by hand and called it WiSE-0.3: Nepali CER down to 10.6% — from one internal pass, not independently re-run, the same evidentiary footing as the soup number above — English essentially unchanged. As far as our own notes show, it's the best teacher we've made — the tower at the top of this page is that checkpoint.
For a while that model lived only in a GPU machine's RAM — if the machine had rebooted before anyone checked the backups, it would have simply stopped existing. We've since fixed the urgent part: it's copied to disk now, and we checked the copy against the original byte for byte. We started a second, off-box copy on a separate machine too, then paused partway through.
The point of a teacher is to make a student small enough to actually ship. We distilled one down to 18M parameters. In English it's fine: first audio in about 41ms, 0.20× real time on one CPU thread, 2.1% Seed-TTS word error rate — matching a separate English-only student built the same way, which also has no Nepali capability at all. In Nepali it fails: 34.7% Seed-TTS CER, more than double the teacher's 16.2% on that same Seed-TTS set (a different scoring protocol than the 10.6%/12.3% "user-suite CER" figures above — the source table doesn't say whether this is the soup teacher or WiSE-0.3 scoring 16.2%, just "teacher"). The source's own words for this result: "fails NE."
The glyphs are Pocket TTS's own vocabulary, Devanagari and Latin, tiled once. The picture underneath is a drawn stencil, not a photograph or a spectrogram — the same storage tower from the key art above, reduced to flat shapes and one lit drive bay so it still reads as a shape once it's made of text.