OpenNMT / OpenNMT/CTranslate2

Whisper using prefix shows no speed up

Open
#1,509 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

question
Dominant language
C++
Stars
4.7k
Forks
536
Avg merge
12h 12m
Merged PRs (30d)
4

Description

Hi, may I know how is the prefix implemented for faster-whisper? I tried looking at the code, it seems like the tokens will be generated as usual (from the start ignoring the prefix) but if the prefix is used than instead of picking the one with the highest probability the prefix token is picked instead.
Original Whisper seems to starting off from where the prefix ends and generating tokens after that.
Not sure if my understanding is correct. Initially thought having prefix will speed up the decoding, but doesnt seems to be the case.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

No file or test is named; start by tracing faster-whisper's prefix handling from its decoding entry point and compare it with the described original Whisper behavior. Done means the implementation's prefix semantics and expected speed characteristics are documented or the reported discrepancy is reproduced with a focused test.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.