Nice. That does seem better (mostly because of the larger file size limit, diarization, word-level timestamps, increased speed, and hallucination and repetition reduction).
I wondered what the rate limits are, looks like:
> Use of Deepgram Whisper Cloud is subject to a rate limit of 50 requests per minute or 15 concurrent requests.
Which is a bit disappointing. I've run into whisper rate limits. That said, Deepgram should be better still because it doesn't have the 25 MB file size limit (2 GB instead).
I feel like OpenAI's Whisper API effort is a bit disappointing - it seems like a half-hearted effort. The problem with a half-hearted effort from the dominant brand is that many/most companies will still choose the half-hearted effort from the brand they know and love, which makes it harder for alternatives to get customers for a better offering, and so customers generally end up worse off.
I think the two good options for OpenAI regarding Whisper API would be either:
1) Offer a Whisper API, and make it really good - lots of iteration, feature additions, service improvements
or 2) Don't offer a Whisper API, tell the public you won't offer one, and encourage other companies to do so, and link out to them / refer users to those other companies
Instead, they're doing 3 - half-hearted effort, which hinders overall offering quality. They could improve on their half-hearted effort either by treating it like an important product, or by pointing users to other companies for when they need more production-level features with Whisper (like they do where they point users to Pinecone and Weaviate).
> Use of Deepgram Whisper Cloud is subject to a rate limit of 50 requests per minute or 15 concurrent requests.
> Which is a bit disappointing. I've run into whisper rate limits.
The rate limit for Deepgram's Nova speech-to-text model is 100 concurrent requests on the pay as you go plan. They claim it's more accurate than their Whisper model but I'm trying out both.
I wondered what the rate limits are, looks like:
> Use of Deepgram Whisper Cloud is subject to a rate limit of 50 requests per minute or 15 concurrent requests.
Which is a bit disappointing. I've run into whisper rate limits. That said, Deepgram should be better still because it doesn't have the 25 MB file size limit (2 GB instead).
I feel like OpenAI's Whisper API effort is a bit disappointing - it seems like a half-hearted effort. The problem with a half-hearted effort from the dominant brand is that many/most companies will still choose the half-hearted effort from the brand they know and love, which makes it harder for alternatives to get customers for a better offering, and so customers generally end up worse off.
I think the two good options for OpenAI regarding Whisper API would be either:
1) Offer a Whisper API, and make it really good - lots of iteration, feature additions, service improvements
or 2) Don't offer a Whisper API, tell the public you won't offer one, and encourage other companies to do so, and link out to them / refer users to those other companies
Instead, they're doing 3 - half-hearted effort, which hinders overall offering quality. They could improve on their half-hearted effort either by treating it like an important product, or by pointing users to other companies for when they need more production-level features with Whisper (like they do where they point users to Pinecone and Weaviate).