The title is dense and the paper is short. But the demo is outstanding: (https://huggingface.co/spaces/aiola/whisper-ner-v1). The sample audio is submitted with "entity labels" set to "football-club, football-player, referee" and WhisperNER returns tags Arsenal and Juventus for the football-club tag. They suggest "personal information" as a tag to try on audio.

Impressive, very impressive. I wonder if it could listen for credit cards or passwords.

I wonder if it could be used to "fire customers" on the customer service line.

"Say 'service' for customer service"

"service"

[TAG: complainer]

<click>

GitHub repo: https://github.com/aiola-lab/whisper-ner

Hugging Face Demo: https://huggingface.co/spaces/aiola/whisper-ner-v1

Pretty good article that focuses on the privacy/security aspect of this — having a single model that does ASR and NER:

https://venturebeat.com/ai/aiola-unveils-open-source-ai-audi...

Wouldnt it be better to run normal Whisper and NER on top of the transcription before streaming a response or writing anything to disk?

What advantage does this offer?

I think one of the biggest advantages is the security/privacy benefit — you can see in the demo that the model can mask entities instead of tagging. This means that instead of transcribing and then scrubbing sensitive info, you can prevent the sensitive info from ever being transcribed. Another potential benefit is in lower latency. The paper doesn't specifically mention latency but it seems to be on par with normal Whisper, so you save all of the time it would normally take to do entity tagging — big deal for real-time applications

Ive worked on some enterprise NER systems (specifically privacy/redaction), and in almost all cases the cost of missing out masking was significantly higher than latency (ofc in an ideal world youd have both).

And in all the research we did, the best solutions ended up passing through a workflow of 1.NN based NER, 2.Regex and 3.Dictionary look ups to really clean information. Using a single method worked well in customer demos but always ended up in what we thought were edge cases in prod.

That being said, latency stuff makes sense. This might work great in conversational use cases. Picking out intent and responding. Every millisecond helps in making things sound natural.

The general principle is that "pipelines" impose a restriction where the errors of the first step get baked-in and can't effectively use the knowledge of the following step(s) to fix them.

So if the first step isn't near-perfect (which ASR isn't) and if there is some information or "world knowledge" in the later step(s) which is helpful in deciding that (which is true with respect to knowledge about named entities and ASR) then you can get better accuracy by having an end-to-end system where you don't attempt to pick just one best option at the system boundary. Also, joint training can be helpful, but that IMHO is less important.

From my experience, ASR-to-NER pipelines don't perform adequately out of the box. Even though SOTA ASR systems claim 85% word accuracy, the distribution of errors is worth looking into. Errors around critical entities like credit card numbers or addresses are particularly prone, and even a small mistake renders the result useless.

These ASR errors cascade into the NER step, further degrading recall and precision. Combining ASR and NER into a joint model or integrated approach can reduce these issues in theory, it's just more complex to implement and less commonly used.

Yeah, I’m also curious about that. Does combining ASR and NER into one model improve performance for either?

Almost definitely. You can think of there being a type of triangle inequality for cascading different systems where manually combined systems almost always perform worse given comparable data and model capacity. Alternatively you have tied the models hands by forcing it to bottleneck through a representation you chose.

Looks like only inference available and no fine tuning code available

Is there any reason why this would work better or is needed compared to taking audio and 1. doing ASR with whisper for instance 2. applying an NER model to the transcribed text?

There are open source NER models that can identify any specified entity type (https://universal-ner.github.io/, https://github.com/urchade/GLiNER). I don't see why this WhisperNER approach would be any better than doing ASR with whisper and then applying one of these NER models.

This works better because it gives a secondary set of conditions for which the decoder (generating text) is conditioning its generation. Assume instead of their demo you are doing Speech2Text for Oncologists. Out of the Box Whisper is terrible because the words are new and rare, especially in YouTube videos. If you just run ASR through it and run NER, it will generate regular words over cancer names. Instead, if you condition generation on topical entities the generation space is constrained and performance will improve. Especially when you can tell the model what all the drug names are because you have a list (https://www.cancerresearchuk.org/about-cancer/treatment/drug...)

On a similar note, I've a request for the HN community. Can anyone recommend a low-latency NER model/service.

I'm building an assistant that gives information on local medical providers that match your criteria. I'm struggling with query expansion and entity recognition. For any incoming query, I would want to NER for medical terms (which are limited in scope and pre-determined), and subsequently where I would do Query rewriting and expansion.

https://www.tonic.ai/products/textual offers NER models through an API or with a UI for managing projects. You can sign up for a free trial at https://textual.tonic.ai

It's so great to see that we finally move away from the thirty year old triple categorization of people, organizations and locations.

This of course means that we now have to think about all the irreconcilable problems of taxonomy, but I'll take that any day over the old version :)

"The model processes audio files and simultaneously applies NER to tag or mask specific types of sensitive information directly within the transcription pipeline. Unlike traditional multi-step systems, which leave data exposed during intermediary processing stages, Whisper-NER eliminates the need for separate ASR and NER tools, reducing vulnerability to breaches."