BuzzASR · decoder latency

Decode latency = token count × time per token

Time per token is fixed by the model, about 19.6 ms, so the only thing that moves latency is how many tokens get decoded. Whisper-ZS and SFT use the same tokenizer and decode the same tokens. FFT trains a tokenizer on Estonian text, so the same sentence takes far fewer.

Rahvusraamatukogu digitaalarhiiv sisaldab tuhandeid vanu ajaleheartikleid.

the national library’s digital archive holds thousands of old newspaper articles · Estonian

Whisper ZS

Whisper tokenizer · zero-shot

0 tok0 ms

SFT

Whisper tokenizer · fine-tuned

0 tok0 ms

same tokenizer, same count

FFT

native Estonian tokenizer

0 tok ✓0 ms

FFT decodes 10 tokens, Whisper 28. At 19.6 ms each: 246 vs 599 ms, 2.4× faster.

token splits from Whisper-large-v3 vs frankenstein_v4/estonian. time per token measured per system; it comes out the same. Estonian FFT CER 1.40.