fix(kokoro): voice loading, Synth method selection, padding fixes

[yocontra]

Fixes for Kokoro TTS native code. Addresses voice data truncation, missing Synthesizer method selection, XNNPACK shape mismatches on repeated inference, progressive speed-up on longer inputs, and phoneme token reordering.

Voice loading reads only 128 of 510 rows

voice_ was a fixed std::array<..., kMaxInputTokens> (128 elements), but hexgrad/Kokoro-82M voice files contain 510 rows. The remaining 382 rows were silently dropped.

Changed voice_ to std::vector, sized dynamically from the file. Also fixed an OOB in voiceID — upstream used std::min(phonemes.size() - 1, noTokens) where noTokens could equal 128, indexing past the end of a 128-element array. Now uses a three-way std::min({phonemes.size() - 1, dpTokens - 1, voice_.size() - 1}).

Synthesizer doesn't do method selection

DurationPredictor discovers and selects from forward_8/forward_32/forward_64/forward_128 based on input size, but Synthesizer only knew about forward. Added the same discovery and selection logic. Falls back to "forward" if no forward_N methods exist, so older models still work.

indices tensor changes size between calls

DurationPredictor::generate() returns an indices vector whose size depends on predicted durations — different per input. XNNPACK caches the execution plan from the first call and errors on shape mismatches.

Fixed by padding indices to context_.inputDurationLimit before passing to the Synthesizer.

Audio progressively speeds up on longer inputs

The Synthesizer's attention mechanism drifts on longer input sequences (60+ tokens), causing later phonemes to be spoken progressively faster than the Duration Predictor intended. The DP's timing predictions are correct, but the Synthesizer compresses later phonemes into fewer samples.

Fixed by capping inputTokensLimit to 60, which forces the Partitioner to split text into shorter chunks that the Synthesizer can render faithfully. Each chunk is roughly one sentence (~15-20 words).

tokenize() scrambles phoneme order on invalid tokens

std::partition was used to filter out invalid (unrecognized) phoneme tokens, but partition does not preserve relative order. When any phonemes fall outside the vocabulary, the remaining valid tokens could be reordered, producing garbled audio.

Changed to std::stable_partition which preserves relative order.

Misc perf

• Skip the durPadded heap alloc (up to 320KB) when DP and Synth use the same token count, which is the common case
• Replace temporary pause zero-vectors with resize() directly on the output
• Move-capture audio in the streaming callback instead of copying

Changes

• Kokoro.h — voice_ from fixed array to vector
• Kokoro.cpp — loadVoice(), synthesize(), generate(), stream(), constructor token limit cap
• DurationPredictor.h — getMethodTokenCount()
• Synthesizer.h — forwardMethods_ member, getMethodTokenCount()
• Synthesizer.cpp — method discovery and selection
• Utils.cpp — stable_partition in tokenize()

[chmjkb]

Thanks for the contribution, @yocontra! We'll be taking a closer look at this soon. In the meantime, if you're open to sharing your experience using this or any other APIs in the library, we'd love to hear your feedback!

[yocontra]

Right now I'm trying to get TTS working well - I was using an onnx based approach and got the quality pretty high (after writing my own phonemizer, because nobody else had theirs quite right). Switched to this library as I need multiple types of things now, and been fixing issues to get the quality to where that other library was.

I'm also sending some PRs to the phonemizer this library uses as there are some words it pronounces wrong and that should get it to perfect again :)

[yocontra]

I'm very eager for #936 as I'm trying to get a fully local replacement for gemini live/chatgpt realtime by doing Speech to Text -> Llama 3.2 -> Kokoro TTS