Transcription fails via OpenAI for long call recordings greater than 25Mb

[stevewebster]

Recently deployed a new 3CX v20 instance for a client and had an issue with call transcription missing for long calls. Wasn't sure why but in the process of trying to manually resubmit to OpenAI for reprocessing it became apparent that OpenAI has a hard 25Mb limit on audio files and in our case we had a 27Mb call recording.

I've seen other threads that v20 removed the option to change the format of call recordings so they are currently pretty big .wav files. We managed to compress this example to a 7Mb .mp3 file and manually resubmit that back to whisper-1 via a powershell script to recreate the transcription.

Not tried Google Speech or built-in 3CX transcription options so not sure if any different.

Has anyone experienced similar and what is the recommended solution to this?

 

[VoIPTools]

I don't know how 3CX has implemented their transcription service, but what I do know is that OpenAI has a 25MB (approximately 30 minutes) file size limit. As a programmer, you can get around this limit by "chunking" larger files into smaller pieces (say 23MB + 2MB) and submit each chunk in separate requests. Then stich the "chunks" (transcriptions) back together and return a transcription that spans all the chunks.

We discovered this OpenAI restriction when we added AI Analysis features to our very popular 3CX Recording Manager tool. In theory, with the use of "chunking" there really is no file size limit. We have transcribed 3-hour phone calls this way. Whether 3CX is chunking is unclear, but if you are seeing limits right at that 25MB file size limit, that would suggest they are not.

To your point, converting the .wav files to .mp3 works just as well as the original .wav file, but has a number of benefits: (1) you can handle longer phone calls without needing to chunk, (2) the file is smaller so requires less bandwidth to upload the recording to AI, and (3) because the file is smaller, it is faster to upload. That is precisely why we do exactly that with our Recording Manager tool. What's not to like? Same accuracy, 75% less bandwidth, and 75% faster upload. That's why we do it

3CX is doing some awesome things with their AI integration. But they are not likely to focus on a deeper dive into our "Ask AI Anything" features, which includes agent scoring (human scoring too), coaching, random sampling, filtering the size and types of calls analyzed, measuring agent performance over time (graphical reporting), instant feedback with agent scorecards, manager notes, granular access rights (eliminate the need for 3CX console access) and much more. Our next release of VoIPTools will include improved diarization (identifying speakers) and also enables you to listen to a recording while following along in the transcription by highlighting the current text.

If you need to take your AI analysis features to the next level, take a closer look at 3CX Recording Manager.

 

[stevewebster]

Does anyone know if this has been fixed in v20 Update 8? Or if we go for Ent+ if the native 3CX transcription also has this limitation?

 

[KyriacosS_3CX]

Does anyone know if this has been fixed in v20 Update 8? Or if we go for Ent+ if the native 3CX transcription also has this limitation?

Hello, a lot of things have been fixed, improved and been added since this thread was originally started. Can you advice what is your exact concern?

 

[stevewebster]

Hi,

In the previous version we had call transcription enabled to use OpenAI API. Any long duration calls (circa 40-50+ minutes) resulted in a .wav file of the recording being > 25Mb in size and therefore transcription failed (https://learn.microsoft.com/en-us/answers/questions/1638419/whisper-model-limitation).

gpt-40 transcribe and gpt40-mini transcribe seem to be even more restrictive with smaller recommended call duration/filesizes.

Assume the only way to resolve would be to chunk the call recordings into smaller file sizes and send for transcription sequentially. But also not sure what AI / Transcription model the 3CX cloud transcription model uses when on ENT+ and if this can handle longer call duration transcription.

Thanks,

Steve

 

[KyriacosS_3CX]

Hi,

3CX Transcription is tested to work for calls up to 3 hours. For other engines it can depend on external variables and can depend on individual setup and limits set.

 

[stevewebster]

Thanks for confirming that. Is there anywhere that we can monitor how many transcription minutes have been used, or current balance, when using ENT+ and 3CX Cloud Transcription? Or is there another way of checking current status?