[Supertonic-3 / LiteRT] Greedy text chunking creates awkward intra-sentence pauses and isolates trailing single words

[ksrvpvl]

Environment

• Model: Supertonic-3-LiteRT
• Component: speech_core::SupertonicTokenizer / LiteRTSupertonicTts
• Target Platform: Android (LiteRT / C++)

---

Description

When synthesizing text using LiteRTSupertonicTts, sentences that slightly exceed the per-chunk token limit (cap / max_codepoints, typically ~50–60 codepoints) suffer from unnatural pauses mid-sentence, frequently isolating the final 1 or 2 words into a standalone synthesis chunk.

This makes the voice sound unnatural, breaking phrases right before the last word (e.g., "... sous les [pause] arbres." or "... dans le [pause] jardin.").

---

Root Cause Analysis

After inspecting supertonic_tokenizer.cpp, the issue appears to stem from the interaction of three mechanisms:

1. Greedy Chunk Packing in SupertonicTokenizer::chunk:
   When an oversize sentence is split on word boundaries, words are greedily accumulated until the capacity (cap) is reached:

   for (char32_t c : sent) {
       if (is_ws(c)) { push_word(); continue; }
       word.push_back(c);
       if (static_cast<int>(word.size()) >= cap) push_word();
   }

   For a typical sentence of ~58–62 codepoints and a cap of ~50, the first chunk consumes ~50 codepoints, leaving a tiny remnant of 7–9 codepoints (almost always exactly the final word) in the second chunk.

2. Forced Terminal Period in SupertonicTokenizer::preprocess:

   if (!ends_with_terminal(out)) out.push_back('.');

   When a sentence is severed mid-phrase (e.g. "... dans le"), preprocess() appends a terminal period .. This forces the non-autoregressive Flow-Matching model to synthesize the fragment with falling terminal intonation, followed by generating "jardin." as a completely new sentence with independent acoustic boundaries.

3. Sub-sentence Punctuation Ignored in kTerm:
   kTerm currently only splits at U".!?…。！？". Clauses separated by commas ,, colons :, or semicolons ; are not treated as valid chunk boundaries, causing long multi-clause sentences to bypass clause-level splitting and fall straight into the greedy word-packer.

---

Steps to Reproduce

Synthesize the following French paragraph with LiteRTSupertonicTts:

Bien sûr ! Voici une petite histoire de vacances d'été en 10 phrases :
Cet été, j'ai passé mes vacances à la campagne. J'ai loué une petite maison près d'un lac magnifique. Chaque matin, je me levais tôt pour faire une longue promenade. J'ai beaucoup nagé dans l'eau fraîche et claire. J'ai aussi passé du temps à lire des livres sous les arbres. Un après-midi, j'ai aidé mes voisins à jardiner. Nous avons mangé des fruits frais cueillis dans le jardin. Le soir, je regardais les étoiles depuis ma terrasse. C'était un été très calme et reposant. Je suis revenu à la ville avec beaucoup de souvenirs heureux.
Est-ce que tu veux que je traduise ce texte ou que je te raconte une autre histoire ?

Observed Behavior

Noticeable pauses and intonation breaks occur at:

• "... de vacances d'été en 10" [pause] "phrases :" (~50 chars + 9 chars)
• "... à lire des livres sous les" [pause] "arbres." (~52 chars + 7 chars)
• "... cueillis dans le" [pause] "jardin." (~50 chars + 7 chars)
• "... beaucoup de souvenirs" [pause] "heureux." (~52 chars + 8 chars)
• "... ce texte ou que je te" [pause] "raconte une autre histoire ?" (~56 chars + 28 chars)

Sentences under ~48 codepoints (e.g., "Cet été, j'ai passé mes vacances à la campagne.") synthesize cleanly without any pauses.

---

[ivan-digital]

Thanks for the precise write-up — the three mechanisms you traced were exactly right, and the fix is in v0.0.13:

• fix(supertonic): stop stranding sentence tails in their own chunk #141 — chunk() no longer word-packs a sentence at the ~56-codepoint budget; a sentence is kept whole and the duration predictor decides (as a preflight) whether it fits the 4.46 s window. Only when it really overflows is it bisected — in balanced halves at the best clause/word boundary via the same splitter Kokoro uses — a fragment that continues gets a trailing , instead of the sentence-final ., and the 0.3 s silence is only inserted where a sentence actually ends. Small overflows (≤10 %) are absorbed by tempo instead of a cut. On your French paragraph: 6 forced splits → 1, 18 inserted silences → 11.
• feat(supertonic): pick a latent-length bucket per piece #144 — the remaining split was structural (the LiteRT graphs are fixed at L=64), so soniqo/Supertonic-3-LiteRT now also ships an L=128 bucket (vector_estimator_L128.tflite + vocoder_L128.tflite, ≈9 s). The host discovers it, loads it on first use, and runs each piece on the smallest window that holds it — your paragraph now has 0 forced splits, every sentence generated in one pass. Re-download the bundle to get the two new files; without them behaviour is the fix(supertonic): stop stranding sentence tails in their own chunk #141 one.

Also fixed on the way: a raw-codepoint cap let NFKD-expanded accented text exceed the 128-token text length and be truncated silently. If you still hear a break somewhere, the piece log line (Supertonic: chunk i/n piece j/m: L=…) tells which window it ran on — please attach it.

[ksrvpvl]

Спасибо Иван, сейчас проверим