Mix real code-switching corpora into training, not just evaluation

[Awshesh12]

All our training data is synthetic. Real code-switched speech is used only for evaluation. The paper's central finding is that this costs a lot of generalisation — the obvious next move is to close the gap from the training side.

Candidate sources

• CS-FLEURS xtts/train — 2097 JA–EN rows already available via data/load_csfleurs.py, currently unused. Note these are TTS, so they help with switch-point modelling but not acoustic realism.
• CS-FLEURS read — real voices, but it is our test set. Would require carving out a disjoint train split and re-freezing the benchmark, which invalidates comparison with the current paper. Do not do this casually.
• SEAME — Mandarin–English, wrong pair, but usable for a cross-lingual transfer experiment
• Japanese monolingual real speech (JSUT, Common Voice ja) mixed with English (LibriSpeech, Common Voice en) — cheap, real acoustics, and the composition ablation in Table 6 already shows monolingual data punches above its weight

Experiment to run

A curve: hold total training volume fixed and vary the synthetic:real ratio (100:0, 75:25, 50:50, 25:75). How much real data is needed before real-speech ScriptAcc saturates? That result would be a genuine contribution on its own and directly extends §8.

Watch out

Any use of CS-FLEURS in training breaks the clean held-out property the current benchmark depends on. Document clearly whichever way it goes.

[Awshesh12]

Partially addressed: real monolingual JA+EN (2,000 FLEURS clips) is now mixed into training as anti-forgetting replay (data/mono/mono_train.jsonl, run_cs_mono.sh) — this cures JA forgetting and improves CS-FLEURS MER. Real code-switched training data remains unavailable (CS-FLEURS read is the only real JA-EN CS set and is reserved for eval), so this issue stays open pending a real CS training source.