[mufradhossain]

The shared transcript estimate (12 tokens/s of audio) is calibrated on Latin-script tokenizers: English realizes ~1.2 tokens/word (~2.6 tok/s of speech), leaving generous headroom. Qwen3-ASR's tokenizer fragments dense scripts — measured 6.0 tokens/word for Bengali and 4.6 for Hindi on identical text — so fast speech in those scripts reaches ~21 tok/s and continuous Bengali dictation truncates at ~25 s (observed: 1307 tokens budgeted for a 105 s clip that needed ~1457).

Give predict_transcript_tokens a per-family rate, default unchanged at 12, and pass 45 from qwen3_asr: the fastest measured speech with 2x margin, still clamped by the decoder context. Validated on a 105 s Bengali clip: rate 12 truncates with a warning, rate 45 decodes the full transcript with no slowdown (13.9 s vs 16.5 s wall clock).

Summary

I was using Bengali dictation through Handy and long clips kept truncating around 25 s. Looking into it: the shared transcript estimate (12 tokens/s of audio) is calibrated for Latin-script tokenizers — running the same text through Qwen3-ASR's tokenizer gives ~6 tokens/word for Bengali vs ~1.2 for English, so continuous Bengali speech outpaces the budget. This gives predict_transcript_tokens() a per-family rate (default stays 12, so no other family changes) and qwen3_asr passes 45.

Scope

• src/transcribe-decode-budget.h — predict_transcript_tokens() gains a tokens_per_sec parameter (default 12)
• src/arch/qwen3_asr/model.cpp — family rate constant + both budget call sites
• tests/decode_budget_unit.cpp — 3 new checks

AI Assistance

I have used GLM 5.3 to implement the C++ changes based on the measurements I found. I reviewed the diff and ran the validations myself.

Validation

• decode_budget_unit: ok — all existing checks unchanged, plus 3 new ones
• End-to-end on a 104.7 s continuous Bengali clip (generated with macOS say -v Piya -r 210, so anyone with a Mac can regenerate it): rate 12 → output truncated at 1307 tokens, exit 1; rate 45 → full transcript through the last word, exit 0, and slightly faster (13.9 s vs 16.5 s — the budget is a ceiling, not an allocation)
• No WER or numerical validation needed: only budget arithmetic changed, and families that keep the default rate are byte-identical (existing tests cover this)

Limitations

• Whisper is not touched: it uses a fixed 256-token cap and a small context, which is a different mechanism and may deserve its own discussion
• Languages other than Bengali were measured at the tokenizer level only