Troubleshooting

The first WhisperX transcription takes a long time

The first time a model is used, it’s downloaded, which can take several minutes depending on your connection and the size of the model (up to ~3 GB). The progress shows when it’s loading. The model stays in memory while its options don’t change, so the next transcriptions start right away.

WhisperX fails with CUDA out of memory

Your GPU doesn’t have enough memory for the settings. Try, in this order:

1. Lower the Batch size (e.g. 4) in Preferences → Transcription.
2. Use a smaller model (e.g. small or base).
3. Use a lighter Compute type (e.g. int8).

The last two can reduce the quality. See Engines for the memory each model needs.

Transcribing takes too long

The speed of WhisperX depends on your hardware, so don’t expect instant results on modest CPUs. Try a smaller model, such as small or large-v3-turbo on a GPU, or the int8 compute type. On a Mac with Apple Silicon, check that Use CPU is off in Preferences → Transcription, so it transcribes on the GPU. Alternatively, use the Whisper API or the Google API, which run on remote servers.

The Whisper API returns the error 429

RateLimitError("Error code: 429 - {'error': {'message': 'You exceeded your current quota, ...', 'type': 'insufficient_quota', ...}}")

Your OpenAI account has run out of credits, or you need to add funds before using the API for the first time (even if you have free credits). Buy credits in the Billing section of your OpenAI account. It may take up to 10 minutes for your account to become active.

If you created the API key before adding funds for the first time and the error persists after 10 minutes, create a new key and set it in Preferences → API keys.

The speakers can’t be identified

If the transcription fails with Identifying speakers requires a Hugging Face token. or Could not download the speaker identification model., the token is missing, isn’t valid or can’t access the model.

Identifying the speakers requires a Hugging Face token, and accepting the conditions of the model. Check that:

• You’ve accepted the conditions of pyannote/speaker-diarization-community-1 with the same account.
• The token has the Read role and is set in Preferences → API keys.

See Identify the speakers.

No microphone is found, or nothing is recorded

• Check that the microphone is connected and click the refresh button next to the list of microphones.
• On macOS, allow Audiotext in System Settings → Privacy & Security → Microphone. On Windows, in Settings → Privacy → Microphone.
• If the level meter shows No sound, choose another microphone in the list or check that it isn’t muted.
• If No audio was recorded. is shown, the recording ended before the microphone sent any sound. Record again, or choose another microphone.

The live text isn’t shown

If The text can’t be shown while recording. is shown while recording, the Live model couldn’t be loaded: for example, it’s downloaded the first time it’s used, which needs an Internet connection, or there isn’t enough memory. The recording isn’t affected, and it’s transcribed as usual when you stop. Choose a smaller Live model (e.g. tiny or base) in the Live text card.

The audio of a transcription can’t be played

The source file was moved or deleted. The text is kept, but the audio can only be played from the original file. The recordings of the microphone and the audio of URLs are kept by Audiotext.

A YouTube video can’t be downloaded

Make sure the URL is correct and the video is public. YouTube changes often, so if it still fails, check for a newer version of Audiotext.

If The YouTube video doesn’t have an audio track. is shown instead, the video has no sound to transcribe.

A link can’t be transcribed

• The URL doesn’t point to an audio or video file.: the link opens a web page, not a file. Only the links of YouTube videos and the direct links to audio or video files work. Look on the page for the link that downloads the file (e.g. the episode of a podcast) and use it, or download the file and transcribe it with the File source.
• The file could not be downloaded: …: the file couldn’t be reached. Check that the link opens in your browser and that you’re connected to the Internet. Links that require logging in can’t be downloaded: download the file yourself and use the File source.

A folder doesn’t transcribe any file

The files that already have a transcription are skipped, and so are hidden files and the ones that match Skip files matching. To transcribe them again, use Re-transcribe all… in the ⋯ menu of the folder, or select them in its list of files. The files of the subfolders are only transcribed with Include subfolders, and the folder must contain supported files.

The Google API asks for the language

The Google API can’t detect the language. Choose the Language of the audio in the settings.

A summary, an answer or a translation fails

• DeepL can’t translate into ….: DeepL doesn’t support that language. Choose another provider, such as a language model.
• The reply of the model was too long., The model didn’t return a valid summary., The model didn’t return a valid reply. or The model didn’t return a valid translation.: the model didn’t write the summary, the answer or the translation in the expected format. Try again, or choose a larger model in Preferences → Providers. The small models of Ollama fail more often.
• For any other error, check that the API key of the provider is set in Preferences → API keys and that your account has credits.

The low-confidence words aren’t underlined

Check that Highlight low-confidence words is on in the ⋯ menu of the transcription. If the transcription has no scores, the menu says why:

• Turn on “Word-level timings” and transcribe it again: WhisperX scores the words when it times them.
• There is no alignment model for its language: the words of that language can’t be timed nor scored.
• Transcribe it again to score its words: it was transcribed before the words were scored.
• Only WhisperX and whisper-1 score the words: the other engines and models don’t.

See Review the low-confidence words.

The video with subtitles shows boxes

If Exported to …, but no font has these characters, so they’re shown as boxes: … is shown after exporting the video with its subtitles, no font of your system can draw those characters. Install a font for that script (e.g. Noto) and export it again.

Updates can’t be checked

Could not check for updates. means that Audiotext couldn’t reach GitHub. Check your Internet connection, or whether a firewall or a proxy blocks it. You can always download the latest version from the releases page.

Something else

Search the issues or ask in the discussions. If you find a bug, report it with your system, the version of Audiotext and the steps to reproduce it.