The Short Answer
Video transcription is the process of converting spoken audio from a video into written text. It's used to create captions, subtitles, searchable transcripts, and content for repurposing. Modern tools use AI speech recognition to do this automatically in seconds — older workflows did it manually, taking hours per video.
This post covers how transcription actually works under the hood, the trade-offs between automatic and manual approaches, accuracy expectations, the major tools, and how After Effects editors integrate transcription into their workflow.
How Video Transcription Works
Under the hood, automatic transcription is a chain of three steps. First, the audio is preprocessed — normalized to a consistent sample rate, often filtered to suppress background noise. Second, a speech-recognition model decodes the audio into a sequence of phonemes and then into word probabilities. Third, a language model picks the most likely word sequence given the phonetic input and the surrounding context.
The big shift in the last few years came from end-to-end transformer models like OpenAI's Whisper and Deepgram's Nova series, which collapsed the old multi-stage pipelines into a single neural network trained on hundreds of thousands of hours of multilingual audio. Output quality jumped significantly, and per-hour cost dropped to fractions of a cent.
Automatic vs Manual Transcription
| | Automatic (AI) | Manual (Human) | |---|---|---| | Speed | 30-90 seconds per 10 min of video | 4-6 hours per hour of audio | | Accuracy on clean English | 92-96% | 99%+ | | Accuracy on noisy / accented audio | 75-90% | 95-99% | | Accuracy on technical jargon | 80-90% | 99%+ | | Cost per hour | $0.01-$0.50 (APIs) or $7/mo (Subflow) | $60-$120 | | Speaker identification | Heuristic, often imperfect | Reliable when humans annotate |
For most video work the right answer is AI first, human review for what matters. The AI does 95% of the labor; the human spends 5 minutes per video catching the words the AI missed, especially proper nouns and brand names.
Use Cases for Video Editors
Captions and subtitles are the obvious one, but transcription has other applications that compound the value:
- Searchable footage libraries. When every clip has an attached transcript, you can grep your archive for the moment someone said "third quarter" instead of scrubbing 40 hours of interview footage.
- Repurposing. A 60-minute podcast episode becomes a blog post, a Twitter thread, and three short video clips — all built off the transcript with minimal new writing.
- Translation pipelines. Translating a transcript is dramatically faster than translating from audio; it also means the translator can work asynchronously without watching the source video.
- Compliance and accessibility. Many platforms and jurisdictions require either captions, transcripts, or both for video content.
- Faster client review. Sending a client a transcript alongside a rough cut lets them mark up exactly which line to trim, instead of writing "remove the part around 2:34".
Major Transcription Tools Compared
| Tool | Type | Accuracy | Pricing | Best for | |---|---|---|---|---| | Whisper (OpenAI) | Open-source model | High | Free (self-host) | Developers, custom pipelines | | AssemblyAI | API | High | $0.12-$0.37 per hour | Custom integrations | | Deepgram | API | High | $0.005-$0.15 per minute | High-volume transcription | | Rev.com | Human + AI | Highest | $1.50/min (human), $0.25/min (AI) | Legal, medical, accuracy-critical | | Descript | Standalone app | Medium-High | $12-$24/mo | Podcast editing, multitrack | | Otter.ai | Standalone app | Medium | $10-$20/mo | Meeting transcription, live | | Subflow | After Effects plugin | High | $7/mo or $70/yr | Video editors working in AE |
The right tool depends on what comes next in your workflow. If you'll edit the transcribed video in After Effects anyway, transcribing inside AE eliminates a round-trip. If you're transcribing meetings to extract action items, Otter is the right tool. If you're transcribing legal depositions, you want Rev's human team.
Using Transcription Inside After Effects with Subflow
Subflow is the After Effects-native option. Instead of running your video through Descript or Rev and importing a caption file back into AE, you transcribe directly inside the comp — the result is one AE text layer per caption cue, time-synced to the audio, in your composition's frame rate.
The workflow:
- Open Window → Extensions → Subflow in After Effects
- Point Subflow at the audio layer of your composition
- Pick a language from the 51 supported (or leave on auto-detect)
- Choose one of three caption modes: Single Word, Smart Flow, or Full Sentence
- Click Generate
Within 30 seconds you have native, editable text layers. Full walkthrough: How to Add Captions in After Effects.
Subflow's output is the AE text layers themselves — it doesn't export to .srt, .vtt, or any other subtitle file format. If you need a separate subtitle file alongside your AE composition (for distribution outside AE, for a translator, for a caption track on a separate player), that's a different toolchain: a dedicated transcription service like Whisper or Deepgram that outputs to .srt directly. Once you have that .srt, the free converter tools at subflow.cc/tools handle format conversions between SRT / VTT / TXT / PDF.
Privacy and Accuracy Considerations
A few things worth weighing when picking a transcription approach:
Where does your audio go? Cloud transcription services upload your audio to their servers for processing. Most reputable providers (Subflow included) don't retain the audio after the transcription pass, but the upload step is unavoidable for cloud-based services. If your audio is highly sensitive — confidential interviews, unreleased product demos, legal recordings — self-hosted Whisper or a regulated provider like Rev is the safer choice.
What's the model trained on? Models trained primarily on English audiobook and podcast data tend to do well on those domains and worse on conversational speech, technical jargon, and accented English. Always test with a representative sample of your actual content before committing to a tool.
How much does post-edit cost? A 95% accurate transcript still has ~50 errors in a 1,000-word transcript. If the use case is internal note-taking, that's fine. If it's broadcast captions, those 50 errors all need to be caught and fixed by a human.
Frequently Asked Questions
What is video transcription, exactly? The process of converting spoken audio from a video into written text. Output is a transcript or a timed caption file. Modern transcription uses AI speech recognition to do this automatically in seconds.
How accurate is automatic video transcription? Top-tier models reach 92-96% word accuracy on clean English audio. Accuracy drops for heavy accents, low-quality recordings, technical jargon, or non-English languages with less training data.
How long does it take? AI transcription is faster than real-time — 10 minutes of video transcribes in 30-90 seconds. Manual transcription takes 4-6 hours per hour of audio. Human-AI hybrid runs about 1 hour per hour of audio.
Is my video uploaded somewhere? Depends on the tool. Cloud services upload your audio for the transcription pass. Some tools run entirely on-device (e.g. self-hosted Whisper). Subflow uses cloud transcription; audio is uploaded for the pass but not stored after.
Difference between transcription and captions? A transcript is the text content. Captions are that transcript broken into timed cues synced to the audio. Most modern tools produce both at once.
What to Do Next
If your captioning workflow currently lives in Descript, CapCut, or Premiere — and most of your video lives in After Effects — moving the transcription step into AE eliminates a round-trip and keeps your typography under your control. See Subflow pricing or install the plugin.