chiprook
← Software
SoftwareOctober 9, 2026, 07:27

Test: faster-whisper int8 dropped up to 60 s of speech, float32 didn't

A developer of a podcast transcription tool found that faster-whisper 1.2.1 in int8 mode with 8 threads skipped entire passages — up to 60.5 seconds in a 61.5-second clip. Gaps appeared in 13 of 15 runs, while float32 produced 0 of 15 at 3.8 s per minute of audio versus 2.4 s for int8.

Test: faster-whisper int8 dropped up to 60 s of speech, float32 didn't
#Faster-whisper#CTranslate2
Read next
Security

CTranslate2 CVE-2026-102566 & CVE-2026-102567: Heap Overflow in AI Model Loader

AI

Cactus releases Whistle speech model at 16.9 MB running on CPU

Software

Ubuntu opens early access to Myna AI dictation

Software

Ubuntu 26.10 'Stonking Stingray' enters beta with more Rust and LLM speech-to-text