Agreed, super bad title.

As an engineer I'm always impressed by a 9X speed-up from state-of-the-art, but I'm guessing that "9X" would make tech-illiterate shareholders unhappy and cause them to short the stock?

Would love a Chrome plugin that "fixes" numbers like this on pages based on crowd-sourced human understanding. People flag incorrect numbers they see on websites; new visitors start to see corrected numbers after enough people flag it with the corrected number.

The same would be useful for e.g. scooters advertising 30 miles range (s/30/10/), laptop battery life of 10 hours (s/10/5/), hardware companies announcing "availability of" a certain product when it's actually not yet for sale (s/availability/unavailability/), etc.

Yeah, I've been wanting a similar chrome extension, actually. HN comments serve as a very valuable filter for information that I learn from, and I think it'd be great if the rest of the web benefitted from that kind of proofreading. Little annotations as a layer on top of the web.

Managing the community would be a bit tough, but if you seeded it with the HN community, and managed to keep the same self-regulating culture, I think that would be a strong base to start from.

> a Chrome plugin

Right, better not support any other browser.

Yeah, DeepSpeech has been notably more accurate from our testing, Kaldi seems to be built and maintained as a research project. From what I've heard, running it at notable scale can be a pain.

Kaldi certainly is a research project. Hopefully in the same vein as Postgres and not so much a toy implementation.

Could you conceivably run DeepSpeech and Kaldi side-by-side on the same audio sample and gain any kind of accuracy benefit?

To start with, errors would have to be uncorrelated, and you would need confidence scores on each word, which I assume you get.

Deepspeech can have multiple models running at the same time so I obviously see a benefit from using another system. Kaldi however is much faster to run in my experience and deepspeech is rnn based which puts a limit on how much you can pararalize it.

There are ways to combine multiple independent speech recognition engines to improve accuracy, e.g. ROVER: http://my.fit.edu/~vkepuska/ece5527/sctk-2.3-rc1/doc/rover/r...

I haven't seen any setups like you describe, but if you want to build a framework for comparing confidence intervals that use case would be possible.

DeepSpeech is definitely not as accurate. You messed something up. Most companies doing speech use kaldi, including Amazon.

Please elaborate on this. How are they using it? What do they do to expand on it?

Indeed these are exactly the numbers they present in Table 2.

I took it as referring to the length of the speech versus the length of transcription time, and not as a comparison against prior technologies.