dhyuk54

dhyuk54 / kg-whisper

Public

Keyword-Guided Whisper with AdaKWS - Paper Reproduction

48
0
100% credibility
Found Apr 04, 2026 at 48 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

An enhanced speech-to-text system that uses keyword spotting and guidance to improve transcription accuracy, especially for specialized audio like medical speech, with an interactive demo interface.

How It Works

1
🔍 Discover KG-Whisper

You hear about a smart tool that turns audio into accurate text by spotting important keywords like medical terms.

2
🚀 Launch the demo

Run a simple command to open a welcoming web page in your browser where everything is ready to try.

3
🎉 Pick your audio

Choose from sample audios like medical talks or best examples, or upload your own file to test.

4
📝 Add keywords

Type in words you care about, like 'heart' or 'pain', so the tool knows what to focus on.

5
▶️ Hit transcribe

Click the button and watch as it shows regular text plus super-accurate version with keywords found and highlighted.

6
📊 Compare results

See side-by-side: plain transcription vs keyword-boosted one, with scores showing big improvements.

✅ Download your transcript

Grab the polished text or full report, perfect for notes, reports, or sharing your clear audio insights.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 48 to 48 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is kg-whisper?

kg-whisper is a Python reproduction of the Keyword-Guided Whisper paper, enhancing OpenAI's Whisper ASR with AdaKWS for open-vocabulary keyword spotting and prompt tuning. It detects domain keywords from audio, then guides transcription for lower WER on noisy or specialized speech like medical dialogues or VoxPopuli. Users get a Gradio demo to compare baseline Whisper, keyword-guided output, and oracle results, plus scripts to train and evaluate on custom datasets.

Why is it gaining traction?

It delivers paper-level gains (e.g., 11% WER on VoxPopuli) without full fine-tuning, just lightweight prompt adaptation and adaptive KWS. The interactive UI lets you upload audio, tweak keyword lists, and batch-eval instantly, bridging research to prototyping. Stands out for handling real-world ASR pitfalls like rare terms in medical or multilingual audio.

Who should use this?

ASR engineers tuning Whisper for verticals like healthcare, where keywords like "myocardial" boost accuracy. ML researchers reproducing keyword-guided papers or experimenting with AdaKWS on custom corpora. Python devs prototyping speech apps needing quick domain adaptation without heavy retraining.

Verdict

Solid repro for keyword-guided Whisper experiments, but 1.0% credibility reflects low stars (48) and early-stage docs—expect some setup tweaks. Grab it if you're hacking ASR prototypes; otherwise, wait for more polish.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.