sacred basin
Audio recorded on Safari does not work through whisper. This is also documented here: https://community.openai.com/t/whisper-api-cannot-read-files-correctly/93420/13
Does anyone have a solve? Current solutions all seem to be to load ffmpeg which seems unideal.
I am having similar issues on firefox, but I have not seen similar comments elsewhere and that might just be a personal issue.
still tide
Hi David, sorry to chime in - I don't have a solution for your problem, but maybe you can help me out with mine, as I am getting a "Bad request" on my calls - how do you send the audio file over via API? Thanks!
gaunt mantle
seems to be broken for me too. according to the api "file=<path to audio>" but I am getting "1 validation error for Request\nbody -> file\n Expected UploadFile, received: <class 'str'> (type=value_error)" so do I send the file's body or the file path?
gaunt mantle
I just managed to post a file successfully. the file parameter is the file data.
sacred basin
@still tide I'm using node, and I'm sending the audio in formdata using an axios post request
formData.append("file", audioStream)``` headers: {
Authorization: `Bearer ${token}`,
"Content-Type": `multipart/form-data; boundary = ${formData._boundary}`,
}
})```But worth noting that I was getting bad request on all my calls with Firefox and safari audio - still havent gotten a solution on that, so if you're trying it out, make sure to be trying it on chrome
still tide
thank you David - I am actually recording on Firefox, I'll follow your advice and try it with Chrome before making changes to the backend
(I am also using Node)
sacred basin
Actually, after seeing your comment I noticed that the openai node package now works for whisper
I just switched over to it - much cleaner!
https://platform.openai.com/docs/api-reference/audio/create?lang=node
their example works without edits
still tide
does it work for you on Chrome?
because I was getting a Bad Request also with their npm package, but I'll now test it with crhome
which mimeType are you using? Maybe that's a topic as well
wait a second
now it works on Chrome
ok, the game is on!
thanks already, let's figure this out
found a way to not have to save it to disk
sacred basin
I got timed-out for my last message but yeah! Mp3, webm, and wav seem to work fine on chrome. The link I have in my top message is active with some people trying to solve it for safari - There is a fix, but tbd on firefox
still tide
const { Readable } = require('stream');
sacred basin
Oh i might try that -i've been just saving and flushing
still tide
Perhaps if the browsers have different internal webm formats, how about converting them on the backend - will check with ffmpeg
sacred basin
yeah someone did that in the support forum - it works but is slow if i remember
still tide
hmmm
not sure, but it looks like the files generated by Firefox are just in the ogg format (just opened the file in the text editor) - that might be an issue
sacred basin
supposedly, but so far it's just broken all my recording logic
still tide
unfortunately the browsers have different formats, and OpenAI seems to be very strict to what they accept
or they don't accept ogg
still tide
I think I was able to make it work in FF with the following setting:
making some more tests - I can't try it on Safari Desktop yet
sacred basin
oh neat - is that using the polyfill media recorder?
sacred basin
Woohoo! That polyfill solution completely fixed my issues, it's now working on safari and firefox and chrome
One caveat - using onstop or ondataavailable don't seem to work - you have to add eventhandlers for them instead (which they loosely cover in their github
still tide
oh neat - is that using the polyfill media recorder?
without the polyfill
still tide
Woohoo! That polyfill solution completely fixed my issues, it's now working on s...
nice!
deep vapor
seems to be broken for me too. according to the api "file=<path to audio>" but I...
I am experiencing the same issue, still tryin to figure it out how can I workaround this:
