Solved Intermittent OpenAI transcription failures – HTTP 400 "Unrecognized file format" with valid WAV files (V20 Update 9 / AI 1.4.48)

[MG_SI]

Hello,

We are experiencing intermittent AI transcription failures.

Environment:

- 3CX V20 Update 9 (Build 995 Release)
- AI Plugin 1.4.48
- AI Edition
- On-Premise Windows
- OpenAI
- Whisper (Speech-to-Text)
- GPT-5 nano (Analysis)

The system worked correctly for several days.

Now transcription works only intermittently.

Symptoms:

- Some calls are transcribed successfully.
- Other calls fail.
- Manual "Transcribe" sometimes does nothing.
- Automatic transcription also occasionally fails.

Event Log sometimes reports:

HTTP 400 (invalid_request_error)

Unrecognized file format.
Supported formats:
flac, m4a, mp3, mp4, mpeg, mpga, oga, ogg, wav, webm

Sometimes we also receive:

Audio file is too short.
Minimum audio length is 0.1 seconds.

Investigation performed:

- OpenAI API key is valid.
- OpenAI credits available.
- AI license active.
- WAV files are valid PCM WAV.
- Both successful and failed recordings have identical technical properties:
- PCM
- Mono
- 8000 Hz
- 16 bit
- Failed recordings play normally in Windows Media Player and VLC.

Interestingly, the downloaded WAV file is valid, but OpenAI reports "Unrecognized file format".

Has anyone experienced this after Update 9 / AI 1.4.48?

Could this be a race condition where AI uploads the recording before it has been fully written to disk?

 

[NikosT_3CX]

@MG_SI does the extension that has the recording, have any special characters in the First Name or Last Name ?
As for the audio length, how long in duration is the recording if you hear it ?

 

[MG_SI]

Thank you for your reply.

Yes, extensions that frequently fail have special characters. Another extension that successfully transcribes recordings has no special characters.

The failed recording duration is 69 seconds. The recording is a valid PCM WAV file (mono, 8000 Hz, 16-bit) and plays correctly in Windows Media Player and VLC.

I will now perform another test after temporarily removing the special character from the user's name and will check the results.

 

[MG_SI]

Thank you, that was indeed the problem.

So it appears that the current AI transcription implementation has an issue handling non-ASCII characters. Thank you for pointing me in the right direction!