ggml-org/whisper.cpp

Specialized vocabulary

Open

#235 opened on Dec 8, 2022

View on GitHub
 (1 comment) (1 reaction) (0 assignees)C++ (5,535 forks)batch import
enhancementhelp wanted

Repository metrics

Stars
 (49,693 stars)
PR merge metrics
 (Avg merge 5d 20h) (36 merged PRs in 30d)

Description

I am interested in using the streaming tool for a very specialized context (radiology dictation), which uses an esoteric and relatively restricted vocabulary (i.e. high likelihood of words like "hyperpneumatization" or "temporooccipital", but never common words like "book" or "spoon"). I have seen references in the whisper documentation to an initial_prompt option to steer the model towards certain terms, but I am not sure if this is feasible to pass in a relatively large corpus of high likelihood, but generally uncommon, words.

In the standard configuration, accuracy for some radiology reports I've tested is fairly poor, preferring more common words over the correct uncommon word. I wonder if anyone has thoughts about this problem.

Contributor guide

Research direction
Examine the whisper.cpp source code to understand how initial prompt is implemented and passed to the model. Test with a large specialized vocabulary (e.g., radiology terms) and measure accuracy. Explore if there are limitations on prompt size and consider alternative approaches such as fine tuning or using a custom vocabulary file.
Tech stack
ccpp
Domain
aiperformance
Issue type
Research
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
Half day
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Clear
Prerequisites
GitC++
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
55