Accurate transcription, ready to read
Turn any recording or live audio into a structured, searchable transcript with speakers, punctuation and formatting in place.
- Smart Formatting
- Speakers, Identified
- Punctuation & Paragraphs
Leading voice teams build with Deepgram
For everything that needs a transcript
Meeting notes, live captions, subtitles, legal records, compliance archives, and lecture capture. All through one API.
Meeting intelligence
Accurate transcripts for meeting notes
Meeting assistants live or die on transcript quality: speaker labels that stay correct across a long call, punctuation that survives cross-talk, and summaries generated from the transcript rather than from the audio a second time.
Real-time accessibility
Real-time captions for live audio
Streaming adds interim results and endpointing, so captions appear as words are spoken and settle as the sentence completes. The same models handle pre-recorded audio, so accuracy doesn't drop just because it's delivered live.
Media and entertainment
Subtitles, archives and search
Word-level timestamps for subtitle timing, diarization for multi-speaker programming, and processing at up to 40x faster than real time, so a back catalogue becomes searchable in hours rather than weeks.
Legal and regulated
Accurate transcripts for legal audio
Depositions, hearings, and recorded interviews need every speaker identified and every word accounted for. Multichannel keeps each speaker on their own recorded line, and redaction removes personal information before you store or share the transcript.
Compliance and retention
Searchable archives at volume
A recorded archive is only useful if you can find things in it. Batch processing turns years of audio into indexed text, with entity detection and redaction applied before anything reaches long-term storage. Run it in Deepgram's cloud or your own environment.
Education and accessibility
Searchable transcripts for lecture audio
Lecture capture, course libraries, and accessibility requirements all need the same thing: an accurate transcript with paragraphs and timestamps, in the language the class was taught in.
Built for volume, priced for it too
The same models behind real-time voice run transcription workloads, so accuracy does not drop when volume goes up.
- >90%word recognition rate (1 - WER)
- 40xfaster than real-time processing
- 2xlower price
Readable, searchable, and ready to use
From formatting and speaker labels to custom vocabulary and structured extras like summaries and redaction, Deepgram turns raw audio into a transcript your team can actually work with.
Smart formatting
Deepgram writes numerals, dates, currency, and addresses the way a person would type them.
Speakers, identified
Diarization labels who spoke when in a single stream, and multichannel keeps each speaker on their own recorded line.
Punctuation and paragraphs
Sentence boundaries and paragraph breaks turn a stream of words into something readable.
Keyterm prompting
Get product names, drug names and industry terms right without retraining a model.
Find and replace
Normalize spellings and correct known variants in the transcript as it’s produced.
Search
Locate spoken terms inside long recordings, returned with timestamps and confidence scores.
Summarization
Generate a summary from the transcript, without sending the audio to a second vendor.
Topics and intent
Detect what a recording is about and what the speaker was trying to do.
Entities and redaction
Identify names, numbers, and identifiers, then remove them before you store the transcript.
FAQs
Transcription produces a record for someone to read, search, or store. A voice agent produces a spoken reply and has to take its turn in a live conversation. Different jobs, different models: Deepgram Nova-3 handles transcription, Deepgram Flux STT handles conversation.
Deepgram Nova-3 handles real-time or pre-recorded transcription, and industry-tuned models cover specialized vocabulary, such as medical or pharma. Deepgram Flux STT is built for voice agents rather than transcription, so it's the wrong choice if your output is a document.
Both, through the same API. Streaming adds interim results and endpointing, so captions appear as words are spoken. Some features are pre-recorded only, including paragraphs, summarization, and topic detection, because they need the whole recording before they can run.
Diarization labels who spoke when within a single audio stream. If you recorded speakers on separate channels, multichannel keeps them apart directly, which is more reliable. Both are available on real-time and pre-recorded audio.
Yes. Redaction removes personal information from the transcript before you store or share it, and entity detection identifies names, numbers, and identifiers so you can decide what to keep. Both run as part of the same request.
You can process a pre-recorded archive at up to 40x faster than real time, so a back catalogue that would take weeks to play through becomes searchable text in hours. Run it in Deepgram's cloud or self-hosted in your own environment.