ASLP-lab

ASLP-lab / Speaker-Reasoner

Public

Speaker-Reasoner: Scaling Interaction Turns and Reasoning Patterns for Timestamped Speaker-Attributed ASR

19
0
100% credibility
Found Apr 09, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
AI Summary

Speaker-Reasoner is a research model that transcribes audio conversations with timestamps, speaker identities, and other details using step-by-step reasoning.

How It Works

1
🔍 Discover Speaker-Reasoner

You hear about a clever tool that turns meeting audio into detailed transcripts showing who spoke when.

2
📖 Explore the idea

You read how it smartly breaks down conversations by first getting the big picture, then zooming into each speaker's part.

3
🏆 Impressive results

You see charts proving it does better than big-name AI tools at spotting speakers and timing words accurately.

4
💻 Prepare your space

You set up a quiet corner on your computer just for handling audio wonders like this.

5
🧠 Get the smart parts

You wait a bit and grab the ready-made thinking pieces to power the tool.

6
🎤 Add your audio

You share your recording of a chat or meeting with the tool.

7
🤔 Watch it reason

It thinks step by step: overview of voices, guessing change points, then detailed notes on each part.

📝 Perfect transcripts

You receive a clear write-up with every speaker named, timed, and transcribed spot-on, even for long talks.

Sign up to see the full architecture

6 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 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 Speaker-Reasoner?

Speaker-Reasoner is a Python-based end-to-end speech LLM for timestamped speaker-attributed ASR, tackling messy multi-speaker conversations by scaling interaction turns and reasoning patterns. It iteratively reasons over audio—summarizing speakers globally, predicting boundaries, then decoding segments with identities, gender, timestamps, and transcripts. A speaker-aware cache lets it handle long-form audio beyond standard context limits without losing consistency.

Why is it gaining traction?

It stands out with agentic multi-turn reasoning that boosts accuracy on benchmarks like AISHELL-4 and AliMeeting, beating baselines including Gemini-2.5-Pro on diarization error rate and concatenated speaker-attributed CER. Developers dig the progressive training for real-world speaker interactions and the promise of SOTA results in timestamped, speaker-reasoner setups without manual segmentation.

Who should use this?

Speech ML engineers building meeting transcription apps, voice agents, or podcast tools needing precise speaker diarization and timestamps. Ideal for teams evaluating ASR upgrades for multi-turn dialogues where off-the-shelf models falter on long interactions.

Verdict

Promising for speaker-attributed ASR with strong benchmark wins, but at 1.0% credibility, 19 stars, and no released models or inference code yet—just a solid paper and setup instructions—hold off until "coming soon" delivers. Watch for ms-swift integration.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.