Engines

Audiotext transcribes with one of three engines, chosen in the Engine card of the transcription settings.

	WhisperX	Whisper API	Google API
Runs on	Your computer	OpenAI servers	Google servers
Internet	Only to download the models	Required	Required
Cost	Free, unlimited	Paid	Free tier (60 min/month), or paid with an API key
Detects the language	✓	✓	✗
Translates	✓	✓	✗
Timestamps	✓	Depends on the model	✗
Identifies speakers	✓ (Hugging Face token)	gpt-4o-transcribe-diarize	✗
Word-level timings	✓	whisper-1	✗
Live text	✓	✗	✗
Low-confidence words	✓ (word-level timings)	whisper-1 (segments)	✗

WhisperX

WhisperX is a fast implementation of OpenAI’s Whisper that runs on your computer, so your audio never leaves it. It runs on the CPU or, much faster, on an NVIDIA GPU with CUDA or on the GPU of a Mac with Apple Silicon.

Model

Larger models are more accurate, but slower and use more memory. The model is downloaded the first time it’s used.

Model	Parameters	Required VRAM
tiny, tiny.en	39 M	~1 GB
base, base.en	74 M	~1 GB
small, small.en	244 M	~2 GB
distil-small.en	166 M	~2 GB
medium, medium.en	769 M	~5 GB
distil-medium.en	394 M	~3 GB
large-v1, large-v2, large-v3	1550 M	<8 GB
large-v3-turbo	809 M	~6 GB
distil-large-v2, distil-large-v3, distil-large-v3.5	756 M	~5 GB

• large-v2 is the default, since large-v3 tends to hallucinate and repeat text more often, especially in some languages like Japanese, and misses more punctuation.
• large-v3-turbo is a pruned version of large-v3, much faster and almost as accurate.
• The models ending in .en (tiny.en, base.en, small.en, medium.en) and the distilled ones (distil-small.en, distil-medium.en, distil-large-v2, distil-large-v3, distil-large-v3.5) only transcribe English. They’re faster than the multilingual models of the same size.

Tip

To try Audiotext quickly, choose tiny or small. For the best quality, use large-v2 or large-v3-turbo on a GPU.

On a Mac

On Macs with Apple Silicon, MLX Whisper transcribes on the GPU, much faster than on the CPU. WhisperX then aligns the words and identifies the speakers on the CPU, as always, so the transcriptions keep their word timings, scores and speakers.

• The MLX version of the model is downloaded the first time it’s used.
• distil-small.en, distil-large-v2 and distil-large-v3.5 have no MLX version, so they run on the CPU.
• The live text always runs on the CPU.
• Compute type and Batch size don’t apply on the GPU, since MLX always calculates in half precision and without batches.
• To transcribe on the CPU, turn on Use CPU in Preferences → Transcription.

Advanced options

They’re in Preferences → Transcription → WhisperX. Change them only if you have problems or know what you’re doing: a GPU that runs out of memory can freeze your system.

• Compute type: the precision of the numbers of the model. float16 is faster on NVIDIA GPUs (the default with CUDA). int8 uses less memory and is the default on the CPU, since many CPUs don’t support float16 efficiently. float32 is the most precise, for GPUs with more than 8 GB of VRAM.
• Batch size: how many parts of the audio are processed at once (8 by default). It doesn’t change the quality, only the speed. Lower it if you run out of memory; up to 16 is recommended.
• Use CPU: runs WhisperX on the CPU. It’s always on if no GPU was found (an NVIDIA GPU with CUDA, or the GPU of a Mac with Apple Silicon).

Whisper API

Uses the speech-to-text API of OpenAI. It’s meant for computers that can’t run WhisperX smoothly, and requires an OpenAI API key (see API keys).

Model	Timestamps	Speakers	Notes
whisper-1 (default)	✓	✗	Can be played segment by segment and subtitled. Translates into English. Scores each segment, to review the low-confidence ones.
gpt-transcribe	✗	✗	More accurate, but without timestamps.
gpt-4o-transcribe-diarize	✓	✓	Identifies the speakers. Doesn’t use the keywords nor the description.

Translations into English are always made by whisper-1, since it’s the only model that translates.

Long audios are split into chunks of up to 10 minutes, cut at a silence so no word is split, since the API rejects files larger than 25 MB. With whisper-1, the end of each chunk is given as context to the next one; with gpt-4o-transcribe-diarize, a sample of each speaker’s voice is sent with the next chunks, so they keep their labels.

Options

• Response format (Output card, for folders): text (default), json, verbose_json, srt or vtt. The subtitles and verbose_json need a model with timestamps.
• Temperature (Preferences → Transcription): between 0 and 1. Higher values like 0.8 make the output more random, and lower values like 0.2 more focused. With 0 (default), the model raises it automatically when needed.
• Timestamps of the words (Preferences → Transcription): whether whisper-1 also returns the timestamps of each word, to highlight it while playing. It takes longer. On by default.

Google API

Uses the Google Speech-to-Text API. It doesn’t punctuate sentences (Audiotext adds the punctuation), and its quality is lower than Whisper’s, so the transcriptions often need corrections. It can’t detect the language nor translate, and returns plain text without timestamps.

Without an API key, the free tier is used, limited to 60 minutes per month. To extend it, set a Google API key. Google charges for its use, for which Audiotext is not responsible.