Offline vs cloud transcription

Uploading a recording hands what was said to another organization. Under the GDPR that organization is a processor: you need a data processing agreement, you need to know where the data is stored and who can access it, and transfers outside the EU need their own safeguards. Many professions add secrecy duties on top, which make sending client or patient audio to a third party something to justify carefully.

Offline transcription removes that whole chain for the transcription itself. The recording is processed where it already is. You still decide how long to keep it and who may read the transcript, but no copy exists anywhere else.

Cloud services usually charge for what you use: per minute of audio for developer services, or a monthly allowance of hours for consumer apps, with the expensive tiers reserved for long meetings and extra features.

Offline, the computing is yours, so the cost per hour can be zero. Transcribe's Standard model is free without limits. Pro, for meeting recording, speakers and the Accurate model without limits, is a flat $49.99 a year: at 20 hours of audio a month that is about $0.21 an hour, and it gets cheaper the more you transcribe.

A cloud service's servers are fast, but your file has to get there first. An hour of uncompressed WAV is over 600 MB, which takes minutes on a slow uplink before transcription even starts.

Offline speed depends on your PC. On an ordinary laptop with integrated graphics we measured Transcribe's Standard model at about 17× real time, an hour of audio in about 3.5 min, and the Accurate model at about 2.5×, an hour in about 24 min. On the processor alone it is slower: about 20 min an hour for Standard. All the measurements.

The best cloud services and the best open models are both very good on clear speech. The differences show up with noise, accents, crosstalk and specialist vocabulary, and they vary by language. No published benchmark predicts how either will do on your recordings.

The honest way to decide is a short test on your own audio: the same hard five minutes through each tool, and a count of the errors in one minute. The ten-minute test.

• A team needs to share and comment on transcripts in one place.
• You want automatic summaries, action items or translation, and the recordings are not sensitive.
• You work mostly on a phone, a tablet or a Mac, where an offline Windows app does not help.
• Your PC is very old and you transcribe many hours under time pressure.

• The recordings are confidential: clients, patients, sources, research participants, internal meetings.
• You work without a reliable connection, or behind strict network rules.
• You transcribe a lot, and per-minute pricing adds up.
• You do not want an account, or to explain a third-party service in a consent form.