You have an interview, a lecture, a podcast episode or a Zoom call, and you need it as text. AI transcription tools do that job, but they split into three groups that suit different people: apps for meetings, apps for editing audio and video, and APIs for developers. A fourth group runs entirely on your own computer.
This guide compares ten tools across those groups, with prices and limits taken from each maker's own page, as of October 2026. Prices change often, so check the linked page before you pay.
Quick picks
- Meeting notes: Otter.ai
- Editing a podcast or video by editing the text: Descript
- Legal or high-stakes work with a human check: Rev
- Captions and subtitles in many languages: ElevenLabs Scribe v2
- Private, on-device transcription on a Mac: MacWhisper
- Free and open source, if you can run Python: OpenAI Whisper
- Open weights with speaker labels and translation: NVIDIA Nemotron 3.5 Transcribe
- Building an app (APIs): Deepgram, AssemblyAI, OpenAI
Apps for meetings and everyday recordings
Otter.ai
Otter.ai joins Zoom, Microsoft Teams and Google Meet calls and writes notes for you. The free Basic plan includes live transcription, speaker identification, 300 monthly transcription minutes and 3 lifetime audio or video file imports. The plan comparison table caps each Basic conversation at 30 minutes of transcription and keeps the 25 most recent conversations. Pro costs $16.99 per user per month on monthly billing and includes 1,200 in-app recording minutes, 10 monthly file imports and meetings of up to 90 minutes. Otter lists AI transcription in English, Spanish, French, German, Japanese and Chinese. The pricing page also lists an Otter MCP server, described as "MCP server integration for AI assistants." Our guide to AI meeting note takers compares Otter with Fathom, Fireflies, Granola and Read AI.
Good for: people who spend the day in video calls. Watch for: the 30-minute cap per conversation on Basic, and the file import caps on Basic and Pro if you mostly upload recordings rather than record calls.
Rev
Rev sells both AI and human transcription, and pitches its subscription plans at legal and investigative teams. The free plan gives 45 AI transcription and caption minutes a month, English only. The Pro plan is listed at $47.99 or $59.99 per seat per month, depending on billing term, with 10,000 AI transcription and caption minutes per user per month and 37+ languages, including Spanglish.
Rev's human transcription costs $1.99 per minute. Rev says it is 99%+ accurate and delivered in 12 hours or less. Legal-formatted transcripts are priced by the page: $2.25 for Legal Rough Draft and $2.50 for Legal Premium Transcript.
Good for: court, research or journalism work where a human check matters. Watch for: human transcription adds up fast on long recordings.
Apps for editing audio and video
Descript
Descript transcribes your recording and lets you edit the media by editing the text. Its pricing page describes it as "Edit video by editing a doc." Transcription covers 25 languages, and the app can detect 8+ speakers and remove filler words such as "um" and "ah."
The Free plan includes 60 media minutes a month. Hobbyist costs $24 per person per month, or $16 on annual billing, and includes 10 media hours a month and 1080p watermark-free export. Creator is $35 per person per month, or $24 on annual billing, with 30 media hours. Descript counts media hours for any media you upload or record, whether you transcribe it or not.
Good for: podcasters and video creators who want transcription and editing in one place. Watch for: media hours, not just transcription time, set your limit.
ElevenLabs Scribe v2
ElevenLabs Scribe v2 is the speech-to-text model behind ElevenLabs Studio and its API. ElevenLabs says it covers 90+ languages. Its speech-to-text page lists speaker labels for up to 32 speakers, word-level timestamps, keyterm prompting for up to 1,000 words or phrases, tags for sound events such as laughter and footsteps, and PII redaction through the API. A separate model, Scribe v2 Realtime, transcribes live speech in under 150 ms, according to ElevenLabs.
On the speech-to-text pricing table, the Free plan gives 10k credits a month, which ElevenLabs lists as 12 minutes of Scribe v2 transcription in the app. Starter costs $6 a month and lists 1 hour 31 minutes. Creator is $22 a month with 6 hours 7 minutes, and Pro is $99 a month with 30 hours 18 minutes. ElevenLabs says every Scribe request passes through its Cue aligner, which sets the timing of each word. We covered it in our ElevenLabs Cue explainer.
Good for: captions and subtitles, especially in languages other than English. Watch for: the free plan's 12 minutes is enough to test, not to work.
Tools that run on your own computer
MacWhisper
MacWhisper is a Mac app that transcribes audio files on your device. Its developer says "All transcription is done on your device, no data leaves your machine," which makes it a fit for sensitive interviews. It supports 100 languages, exports .srt and .vtt subtitles as well as docx, pdf and other formats, and can record meetings in Zoom, Teams and other apps. It needs macOS 14 or later.
MacWhisper Pro is a one-time payment, not a subscription. Pro adds NVIDIA Parakeet models, speaker diarization with NVIDIA Nemotron 3 Diarization, batch transcription and YouTube transcription. Pro can also send audio to cloud transcription services such as OpenAI, ElevenLabs and Deepgram, so keep to the local models if the audio must stay on your Mac. The developer notes that the Medium and Large Whisper models need a Mac with more than 8GB of RAM.
Good for: journalists, researchers and anyone who cannot upload audio to a cloud service. Watch for: the desktop app is Mac only. An iPhone and iPad version is a separate download.
OpenAI Whisper (open source)
Whisper is OpenAI's open speech recognition model. The README says "Whisper's code and model weights are released under the MIT License." It handles multilingual transcription, translation into English and language identification. The default turbo model has 809M parameters and needs about 6 GB of VRAM. The README warns that turbo is not trained for translation, so use one of the multilingual models, such as medium or large, to translate non-English speech into English.
Good for: developers and tinkerers who want a free tool with no usage limits. Watch for: you install it with Python and run it from the command line. Apps such as MacWhisper wrap it in a normal interface.
If you build Mac or iPhone apps, our report on AnySpeechRecognizer covers a Swift package that runs Parakeet and Qwen3-ASR behind Apple's speech API.
NVIDIA Nemotron 3.5 Transcribe (open weights)
Nemotron 3.5 Transcribe is a newer open-weight model from NVIDIA. Its model card lists transcription in 40 languages, recordings up to one hour, speaker-attributed transcription, speech translation and timestamp prediction. It has 34B total parameters, with about 4.3B active, and is released under the OpenMDW-1.1 license. NVIDIA's serving figures use an RTX PRO 6000 Blackwell Server Edition GPU, so it suits teams that host their own models. Our Nemotron 3.5 Transcribe post has the benchmarks.
Good for: teams that want speaker labels and translation without sending audio to a vendor. Watch for: the hardware it needs.
APIs for developers
If you are building an app, a call center tool or a voice agent, you pay per minute or per hour of audio. The rates below are each provider's listed standard rates for recorded (not live) audio.
Deepgram
Deepgram charges $0.0043 per minute for Nova-3 Monolingual on recorded audio, which works out to about $0.26 an hour. Live streaming with Nova-3 Monolingual is $0.0077 per minute. Deepgram says Nova models support 60+ languages, bills per second with no rounding up, and includes speaker diarization at no extra cost on recorded audio. New accounts get a $200 credit, which Deepgram says covers roughly 43,000 minutes of Nova-3 transcription.
AssemblyAI
AssemblyAI charges $0.21 an hour for Universal-3.5 Pro and $0.15 an hour for Universal-2 on recorded audio. Its live model, Universal-3.6 Pro Realtime, is $0.45 an hour. Add-ons stack on top: standard speaker diarization (who spoke when) is +$0.02 an hour on recorded audio, and Medical Mode is +$0.15 an hour. New accounts get $50 in free credits. One detail to note: AssemblyAI bills live streaming on how long the connection stays open, not on how much audio you send.
OpenAI
The OpenAI API lists gpt-transcribe at an estimated $0.0045 per minute, or about $0.27 an hour. gpt-4o-mini-transcribe is about $0.003 per minute ($0.18 an hour), and gpt-4o-transcribe-diarize, which adds speaker labels, is about $0.006 per minute. The hosted Whisper model is $0.006 per minute. For live audio, gpt-live-transcribe is about $0.017 per minute.
Here are the recorded-audio rates side by side, converted to an hourly figure:
- AssemblyAI Universal-2: $0.15 per hour
- OpenAI gpt-4o-mini-transcribe: about $0.18 per hour ($0.003 per minute)
- AssemblyAI Universal-3.5 Pro: $0.21 per hour
- Deepgram Nova-3 Monolingual: about $0.26 per hour ($0.0043 per minute)
- OpenAI gpt-transcribe: about $0.27 per hour ($0.0045 per minute)
Price is only part of the choice. Test each API on your own audio, with your own accents, background noise and jargon, before you commit. Google also offers a transcription model, Gemini 3.5 Transcribe; see our Gemini 3.5 Transcribe post.
How to choose
Ask four questions:
- Where does the audio come from? Live calls point to Otter.ai or a streaming API. Uploaded files work with any tool here.
- What happens after the transcript? If you edit the audio or video next, Descript saves a step. If you need subtitles, look at ElevenLabs Scribe v2 or MacWhisper's .srt and .vtt export.
- Can the audio leave your machine? If not, use MacWhisper, Whisper or Nemotron 3.5 Transcribe.
- Do you need a guarantee? Rev sells human transcription with a stated 99%+ accuracy. ElevenLabs also sells human transcription through its Productions team, from $2.00 per minute of audio, in a beta open to select users.
Tips for better transcripts
- Record each speaker on a separate track if you can. Descript lists multitrack transcription with speaker labels per track.
- Add names and jargon to a custom vocabulary. Otter.ai, ElevenLabs, Deepgram and AssemblyAI list a vocabulary or keyterm feature on their pricing pages, and Descript has a transcription glossary of up to 30 terms, in English only.
- Check numbers, names and quotes by ear before you publish anything.
- For music, separate the vocals first. Our guide to AI stem splitters and vocal removers covers the tools.
For voice generation and music app prices, see our AI voice and music pricing comparison.
Sources
- Otter.ai pricing
- Rev pricing
- Descript pricing
- Descript Help: Transcription glossary
- ElevenLabs Speech to Text (Scribe v2)
- ElevenLabs docs: Transcripts (human transcription)
- MacWhisper on Gumroad
- OpenAI Whisper on GitHub
- NVIDIA Nemotron 3.5 Transcribe model card
- Deepgram pricing
- AssemblyAI pricing
- OpenAI API pricing
