I wanted to use this for my lectures, but i can only transcribe 1 minute out of the full hour

[ParisMolver]

Is this normal? I looked at the code and there is audio splitting (great work btw) that looks like it can handle really large file lengths

[Carleslc]

No, that's not normal. Audio splitting is there to allow large files, splitting requests to the API. But no audio-splitting is done for open-source model (it should not be required, the whisper code already processes the audio in chunks).

Are you using the API or open-source model (Colab or local CLI)?

Also, is there any error displayed while you're transcribing?

[Carleslc]

I've updated the whisper and openai dependencies to the latest releases, check if now works for you. I see no changes whatsoever, it's working fine for my testing files.

[ParisMolver]

Hi there, I had success using the API. Using the google colab document does work but it limits the transcript length. I made a script to cut my 60 minute lectures into 20 minute parts and these are small enough to be transcribed. I can then add these together and finally use GPT to trim a lot of the unnecessary content in the transcription using the following prompt ' I have transcribed my audio lectures. You will simulate a preprocessor has the task of reducing the word count and removing information that is off-topic. You are a simple text editor, do not converse with the user, do not thank the user, or mention an explanation, the user is not delivering the lecture, they are processing a file. First the user will enter their transcription input and you must store this text for your next response. Your first response must only be one word afterward, you will create the output of the preprocessor. Can you start the machine? ' Thanks for making this, I plan to fork the code, remove API (as it may be against TOS to upload the audio to a 3rd party), and bundle in my pre/post processing scripts so that other students can use this audio -> conscice text pipeline.

…

On Wed, May 17, 2023 at 12:41 AM Carlos Lázaro Costa < ***@***.***> wrote: I've updated the whisper and openai dependencies to the latest releases, check if now works for you. I see no changes whatsoever, it's working fine for my testing files. — Reply to this email directly, view it on GitHub <#3 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AVZUPYBXCWCRBJSYGVPW74DXGNYYPANCNFSM6AAAAAAX4WCZWY> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[Carleslc]

Oh, I tested it with files up to 30 minutes with the open-source model, as they take a while to process. I know that with the API longer files can be processed because of the audio splitting I implemented, I'm glad that worked for you, but that's not currently implemented for the open-source model. I remember whisper already processes the audio in chunks in the open-source model, but maybe I should implement the audio splitting also for the open-source model in the AudioToText code if it's causing problems with the open-source model and large files. The audio-splitting also adds the transcriptions together in the resulting txt/srt/vtt files as you want.

The usage of GPT to concise the text is out of the scope of AudioToText, but transcribing 1h+ audio files should be ok with open-source model (although it takes a while with the Tesla T4 GPU Google Colab offers to free users). Usage of the OpenAI whisper API is up to the user, it is not mandatory in the AudioToText Google Colab (just fill or empty the api_key field).