@_aiOla I just saw your Jargonic announcement.
Your post is not really speaking to developers. Programmers won't schedule a call with you just to try your API.
I'd like to test your ASR to see if it's suitable for production use at Happy Scribe. Can you DM the details?
@flavioschneider Quick update: I did a vibe-check with 11 audio files in 4 languages and I am very very impressed.
Great job!
(still looking forward to more detail if you can provide it!)
@flavioschneider Hey Flavio! Yoel from Happy Scribe here :)
I'd love a blogpost with much more technical detail and granularity about the quality of the model. (diarization metrics for example)
Are you planning on writing one?
Otherwise no worries, we'll run it through our internal benchmarks!
@YesItsRajat@dvassallo Fast forward to today, we have Kamal: https://t.co/9UPBljXeeo
We've used it to migrate some unnecessary microservices back home to our Majestic Monolith.
If you are curious about how it looks: https://t.co/guwgHPjPJI
You also have good-old https://t.co/LSdm6S5YGo, since 2009
@deepgramscott@DeepgramAI@happy_scribe BTW here are the WER in our benchmarks - using whisper's normalizer:
happy_whisper: 0.08 (our finetuned version of whisper)
assembly_ai: 0.08
speechmatics: 0.1
rev: 0.11
deepgram_nova: 0.13
google: 0.32
Possible unfair advantage: our references are clean read, nova is verbatim
@deepgramscott@DeepgramAI@happy_scribe Awesome, thanks for the reply @deepgramscott 🙌 I am in touch with your team!
It's quite a varied benchmark with ~20H of audio and ~100 files. Not focusing on any specific domain, but with a higher-than-average difficulty.
The references follow: https://t.co/6FhjTqCY6C
@karpathy "Eval is hard", yes! That is the main criticism I have about the paper. As we get closer to Human performance, WER will be capturing more and more noise. It's not a good metric to evaluate Humans.
https://t.co/r8Y4E72Xin
We'll test with real transcribers at @happy_scribe
@OpenAI@OpenAI I would like to challenge the claim that Whisper is close to human-level accuracy. To me, this graph suggests that WER might not be the right metric to compare with humans at this level of accuracy. Transcription is hard, but not so hard for humans to get 8% of it wrong.
@OpenAI Still, I've been playing around with Whisper, and it is really, really impressive. Specially if you look at the performance on minority languages such as Catalan or Galician.
Thanks for pushing the state of the art forward 🦾
@OpenAI@OpenAI I would like to challenge the claim that Whisper is close to human-level accuracy. To me, this graph suggests that WER might not be the right metric to compare with humans at this level of accuracy. Transcription is hard, but not so hard for humans to get 8% of it wrong.
🚀 Time is flying and the EU-Startups Summit is coming up fast!🚀We're so excited to welcome you to Barcelona on May 12-13 for 2-days of unmissable startup action 🥳 Joining us will be:
🌟 Happy Scribe Co-Founder and CTO Marc Assens Reina 🌟
https://t.co/m4189KZh4P
@Elena20Ruiz :facepalm: esto se llama alquiler de temporada y hay portales immobiliarios que lo ofertan en una sección a parte (no es el caso de idealista, pero habitaclia sí por ejemplo). Puede estar bien para estudiantes por ejemplo, porque normalmente es más barato. Pero WTF 975 euros
@joelgascoigne It has worked for me so far! My company started a similar thing on April for those of us who have kids, and it has worked wonders for me and my family. Go for it!