Rev AI

Batch jobs handle pre-recorded audio and video, and Rev says hour-long files come back in under a minute. The practical ceiling is 17 hours of audio per job, and files sent by URL can reach 5 TB. That covers full day recordings, depositions and conference archives without splitting them first.

The streaming API turns live audio into captions with low latency. It runs on the same models as the batch product. It supports WebSocket for apps and RTMP for broadcast encoders. Streams run up to 3 hours, so long events need a reconnect step that Rev documents.

Rev AI labels who said what, with support for up to 8 speakers in English and 6 in other languages. If each person was recorded on a separate track, you can pass the channel count and get cleaner separation than diarization alone. No setting exists to force a speaker count, which matters on messy group recordings.

Every transcript can carry timestamps down to the word, which makes quoting, search and caption sync far easier. Forced alignment is the related feature: feed it audio plus an existing script and it returns exact timing per word. Media teams use it to sync burned-in subtitles, and legal teams use it to cite a moment in a deposition.

The asynchronous API supports 58 languages, including multilingual English and Spanish models, while streaming covers 9. Language identification runs as a separate call, so you can sort unknown audio before spending money on transcription. Each language has its own accuracy profile, and Rev publishes the current list.

Beyond text, the audio transcription service can score sentiment, extract topics, summarize a long recording and translate into another language. Each is a separate parameter on the job, which keeps you from paying for features you skip. Standard and premium models are priced differently, so check the rate before running a large batch.

You can supply a list of names, products and technical terms so the model spells them the way your business does. This helps most with brand names, drug names and internal jargon that general models mangle. It's a plain word list, not a language model you train.

Official SDKs cover Python and Node.js, webhooks signal job completion, and output formats include JSON, plain text, SRT and VTT. Rev also offers self-hosted deployment for buyers who can't use a public cloud. Jobs stay on the server for 30 days unless you shorten that window.