While working with the impressive Speech-to-Text language model Whisper from OpenAI, we came upon a slick little app that allows you to easily selfhost Whisper with NodeJs. It's well designed and works great. So we installed it here at whisper.tensorvoice.org and are making it available for you to use freely.
We know there are a lot of these Whisper endpoints popping up all over the Internet. The best one yet is the one that runs entirely inside your browser using Web Assembly, but, alas, it's not very fast. Cool but not yet practical. You really do need a server for this kind of processing.
If you know anything about TensorVoice, you know we're not big fans of sending private data into cloud servers. Our whole business model is oriented around keeping your data secure by keeping it on your premises. So although, yes, we're running this "in the cloud" we're making this Whisper endpoint available with our usual commitment to privacy: we keep no logs of audio or text transcripts. We also do not use cookies1.
Not everyone makes that kind of commitment to privacy, so there is a small potential that a lot of people will use this if word gets out. Until we get a sense of traffic flow, we're limiting the ability to transcribe to 10 minutes to ensure our server doesn't get overwhelmed. As this is a useful length for certain kinds of testing that we do, we'll occasionally use the same endpoint as you for some of our internal research.
Have fun!
Footnotes:
1. Note, although we keep none of your data, we do understand basic cybersecurity and therefore log IP addresses, simply as a matter of normal Apache webserver logging. We currently do nothing with that data but use it in aggregate to notice trends like: if a lot of you start using this whisper endpoint, then we will migrate it from a CPU to a GPU to make it faster. We purge such logs every year or so.