Ask HN: What Speaker Diarization tools should I look into?

11•justforfunhere•6mo ago

Hi,

I am making a tool that needs to analyze a conversation (non-English) between two people. The conversation is provided to me in audio format. I am currently using OpenAI Whisper to transcribe and feed the transcription to ChatGPT-4o model through the API for analysis.

So far, it's doing a fair job. Sometimes, though, reading the transcription, I find it hard to figure out which speaker is speaking what. I have to listen to the audio to figure it out. I am wondering if ChatGPT-4o would also sometimes find it hard to follow the conversation from the transcription. I think that adding a speaker diarization step might make the transcription easier to understand and analyze.

I am looking for Speaker Diarization tools that I can use. I have tried using pyannote speaker-diarization-3.1, but I find it does not work very well. What are some other options that I can look at?

Comments

nemima•6mo ago

Hi, I'm an engineer at Speechmatics. Our speech-to-text software handles speaker diarization very reliably, and we're a go-to choice for non-English languages. https://www.speechmatics.com/

How long is the audio file? If it's under 2 hours, you can upload the file and transcribe it with diarization for free using our web portal: https://portal.speechmatics.com/jobs/create/batch

Hope it helps for your use case! If it does, and you encounter any issues, drop us an email at devrel@speechmatics.com :)

EDIT: typo

justforfunhere•6mo ago

Hi, yes, it is well under two hours. The longest audio that I have had to handle as of now is around 10 minutes.

I will give your portal a try soon. Thanks

hildekominskia•6mo ago

Skip pyannote 3.1; two battle-tested upgrades:

1. NVIDIA NeMo’s `diar_msdd_telephonic` (8 kHz) or `diar_msdd_mic` (16 kHz) — one-line Python install, GPU optional, beats pyannote on cross-talk. 2. AssemblyAI’s async `/v2/transcript` endpoint — gives you `words[].speaker` + Whisper-level accuracy for 40+ languages. Free tier: 3 h / month.

Glue either to your existing Whisper pipeline and feed ChatGPT-4o with speaker-tagged text. The jump in clarity is night-and-day.

I use the same combo to auto-caption interviews, then drop the synced footage into Veo 3 (https://veo-3.app) for instant talking-head explainers—works even for non-English audio.

hbredin•6mo ago

Hey, I am the creator of pyannote open-source toolkit.

I just created a company around it that serves much better diarization models through an API.

You can test it by creating an account on https://dashboard.pyannote.ai. You'll get 150h of diarization for free.

There is also a playground where you can simply upload a file and visualize the diarization results.

satvikpendem•6mo ago

Seems like this only diarizes, is there a transcription interface as well? The prices are a bit high for only diarization as something like Soniox is also ~13 cents for real-time diarization with transcription included.

satvikpendem•6mo ago

Google Gemini and ElevenLabs are quite good at transcription with diarization if you already have the audiofile. For real-time, I like Soniox, you can use their comparison page that runs all the major transcription services at once [0]. Note that their Google model is not Gemini, it's their older Chirp model.

[0] https://soniox.com/compare/

vismit2000•6mo ago

Elevenlabs does speaker diarization really well in my experience: https://elevenlabs.io/ (First came to know about this from Lex-Modi podcast)

meerab•6mo ago

I am building VideoToBe.com - I have found that whisperX works the most reliable.

https://github.com/m-bain/whisperX

It is built on top of OpenAI Whisper, so speech recognition is good, the transcript gives speaker tags as 'SPEAKER_00' and 'SPEAKER_01' etc.

Here is how the transcript may look like

https://videotobe.com/play/media/1b02f75a-9503-43aa-8956-d18...

Hoot: Scheme on WebAssembly

What the longevity experts don't tell you

Monzo wrongly denied refunds to fraud and scam victims

They were drawn to Korea with dreams of K-pop stardom – but then let down

Show HN: AI-Powered Merchant Intelligence

Bash parallel tasks and error handling

Let's compile Quake like it's 1997

Reverse Engineering Medium.com's Editor: How Copy, Paste, and Images Work

Go 1.22, SQLite, and Next.js: The "Boring" Back End

Laibach the Whistleblowers [video]

Slop News - HN front page right now hallucinated as 100% AI SLOP

Economists vs. Technologists on AI

Life at the Edge

RISC-V Vector Primer

Show HN: Invoxo – Invoicing with automatic EU VAT for cross-border services

A Tale of Two Standards, POSIX and Win32 (2005)

Ask HN: Is the Downfall of SaaS Started?

Flirt: The Native Backend

OpenAI's Latest Platform Targets Enterprise Customers

Goldman Sachs taps Anthropic's Claude to automate accounting, compliance roles

Ai.com bought by Crypto.com founder for $70M in biggest-ever website name deal

Big Tech's AI Push Is Costing More Than the Moon Landing

The AI boom is causing shortages everywhere else

Suno, AI Music, and the Bad Future [video]

Ask HN: How are researchers using AlphaFold in 2026?

Running the "Reflections on Trusting Trust" Compiler

Watermark API – $0.01/image, 10x cheaper than Cloudinary

Now send your marketing campaigns directly from ChatGPT

Queueing Theory v2: DORA metrics, queue-of-queues, chi-alpha-beta-sigma notation

Show HN: Hibana – choreography-first protocol safety for Rust