fix(parakeet): clamp multilingual v3 chunk windows for non-English input

[nachoal]

Context

Parakeet v3 multilingual decoding becomes unstable on oversized windows. In practice, non-English audio can drift into English/Spanglish when chunk_duration is large.

This PR keeps decoder behavior unchanged and fixes the issue at the chunk scheduler level.

Description

This change adds a small scheduling guard for multilingual Parakeet v3:

• Detect multilingual Parakeet v3 from the model vocabulary.
• Accept an optional language hint in the Parakeet generate/decode path.
• When the model is multilingual v3 and the caller provides a non-English language hint, clamp oversized windows to:
  • chunk_duration = 5.0
  • overlap_duration = 1.0
• Leave English / auto behavior unchanged.

The goal is to preserve language consistency on non-English input without introducing unsupported decoder prompting hacks.

Changes in the codebase

• Add multilingual Parakeet v3 detection helper.
• Add non-English language-hint helper.
• Add _resolve_multilingual_chunking(...) in the Parakeet base model.
• Thread the optional language hint through the Parakeet model methods so it can be used strictly for scheduling.
• Add STT model tests for:
  • v3 multilingual detection
  • clamping on non-English hints
  • preserving baseline behavior for English / auto

Changes outside the codebase

No infrastructure or service changes.

Additional information

Targeted repo tests:

python3 -m pytest -q mlx_audio/stt/tests/test_models.py -k 'multilingual_chunking or multilingual_v3_detection'

Result:

• 3 passed

End-to-end benchmark summary on a local stt wrapper using this branch:

• English short: unchanged (1123 words)
• English long: unchanged (1241 words)
• Spanish short: fixed (583 -> 1123 words vs prior bad MLX behavior)
• Spanish long: preserved quality, with some throughput cost due to shorter windows

Subtitle-backed Spanish validation (WER):

• es_short: 0.0758
• es_long: 0.1119
• es_long_motivar_humor: 0.0768
• es_long_liderazgo_inspira: 0.0708
• es_long_productividad: 0.0528
• es_long_atencion_inteligencia: 0.0612
• es_long_cerebro_atento: 0.0847

The tradeoff is that non-English Parakeet v3 gets slower than the large-window path, but the prior behavior could silently return the wrong language.

Checklist

• Tests added/updated
• Documentation updated
• Issue referenced (no existing repo issue for this specific fix)

[nachoal]

Follow-up validation summary.

Targeted repo tests:

python3 -m pytest -q mlx_audio/stt/tests/test_models.py -k 'multilingual_chunking or multilingual_v3_detection'

Result:

• 3 passed

End-to-end benchmark on the original sample set:

Sample	Backend	Time (s)	RTF	Words
en_short	mlx-parakeet	6.65	0.0137	1123
en_long	mlx-parakeet	6.80	0.0122	1241
es_short	mlx-parakeet	7.55	0.0187	1123
es_long	mlx-parakeet	15.57	0.0155	2060

Reference-backed Spanish validation (subtitle/text reference WER):

Sample	WER
es_short	0.0758
es_long	0.1119
es_long_motivar_humor	0.0768
es_long_liderazgo_inspira	0.0708
es_long_productividad	0.0528
es_long_atencion_inteligencia	0.0612
es_long_cerebro_atento	0.0847

Interpretation:

• English stays at baseline quality.
• The known bad short Spanish case is fixed.
• The broader long-Spanish validation set stays in a usable WER band while avoiding the previous language drift failure mode.
• Tradeoff: non-English Parakeet v3 gets slower because it now uses smaller windows when a non-English language hint is provided.

[lucasnewman]

@nachoal Can you retest with the fixes from #592? It should be more faithful to the original implementation now.