The meeting transcription market is splitting into two distinct architectures: local AI that runs on your device, and cloud AI that processes audio on remote servers. Both use the same underlying technology (OpenAI's Whisper is the dominant model in both camps), but the implementation creates fundamentally different user experiences.
This guide explains the technical and practical differences so you can choose the right approach for your needs.
How each approach works
Local AI (on-device transcription)
Local AI downloads a machine learning model to your computer and runs it there. The entire pipeline — audio capture, speech recognition, text formatting — happens on your hardware.
The stack looks like this:
- Audio capture: Your computer's microphone and system audio output
- Pre-processing: Voice Activity Detection (VAD) filters out silence
- Inference: The Whisper model converts audio to text using your CPU or GPU
- Post-processing: Punctuation, speaker labels, and formatting are applied locally
- Storage: Transcripts are saved to your local disk (usually SQLite or JSON)
Tools that use this approach: Clearminutes, Whisper.cpp, Jamie (partial), Granola (partial).
Cloud AI (server-based transcription)
Cloud AI uploads your audio to a vendor's server, where it's processed on powerful hardware and the text is sent back to you.
The stack looks like this:
- Audio capture: Your computer or a meeting bot records the audio
- Upload: The audio file is transferred to the vendor's servers (AWS, Google Cloud, Azure)
- Inference: The vendor runs Whisper (or their own model) on GPU clusters
- Post-processing: Summaries, action items, and analytics are generated server-side
- Delivery: Results are sent back to your device and stored in the vendor's database
Tools that use this approach: Otter.ai, Fireflies.ai, Fathom, tl;dv, Krisp (transcription only).
Accuracy comparison
Here's the surprising truth: local and cloud transcription use the same model. Otter, Fireflies, and Clearminutes all run Whisper under the hood. The accuracy difference comes from model size and audio quality, not architecture.
| Factor | Local AI | Cloud AI |
|---|---|---|
| Model used | Whisper (your choice of size) | Whisper (vendor's choice) |
| Typical WER | 3-6% depending on model size | 3-6% depending on vendor |
| Custom vocabulary | Possible with fine-tuning | Usually not available |
| Audio preprocessing | Depends on your setup | Usually better (noise gates, normalisation) |
| Speaker labels | Available (diarization models) | Available (vendor-specific) |
The bottom line: for clean audio with native English speakers, both approaches are effectively identical. For noisy audio, accented speech, or technical jargon, cloud vendors usually have better preprocessing pipelines — but you can achieve the same results locally with a good microphone and some audio engineering.
Speed: real-time vs batch
Cloud AI has a structural advantage for real-time transcription. Server GPUs are more powerful than laptop chips, and cloud services can process audio faster than real-time even with the largest Whisper models.
However, the "real-time" advantage is often overstated. Here's why:
- Upload latency: A 30-minute meeting recording is roughly 30 MB. Uploading this takes 10-30 seconds on a typical connection. The "instant" cloud transcription isn't instant — it's instant after the upload finishes.
- Local GPU acceleration: On an M3 MacBook Pro or a laptop with an NVIDIA GPU, Whisper large-v3 processes audio at roughly 1x real-time. A one-hour meeting takes about an hour to transcribe. The medium model runs at 2-3x real-time.
- Streaming: Local tools can stream transcription in real-time as the meeting happens, just like cloud tools. The delay is milliseconds, not minutes.
For batch processing (transcribing after the meeting), cloud is faster if you have good upload bandwidth. For real-time streaming during the meeting, local and cloud are comparable.
Cost over time
This is where local AI has a clear advantage for heavy users:
| Usage pattern | Local AI cost | Cloud AI cost |
|---|---|---|
| 5 hours/week, personal use | £0 (free tiers available) | £0-15/month |
| 20 hours/week, professional | £0-12/month (Pro for features) | £15-30/month |
| Team of 10, 40 hours/week each | £0-120/month | £150-400/month |
| Enterprise (100+ users) | £0-1,200/month | £1,500-5,000/month |
Local AI is essentially free after setup. The only ongoing cost is electricity (negligible) and optional Pro subscriptions for features like cloud sync or team sharing. Cloud AI charges per minute, per user, every month, forever.
For a consultant transcribing 20 hours/week, the five-year cost difference is approximately £3,000 in favour of local AI.
Privacy architecture
This is the most significant difference between the two approaches, and it's not close.
Local AI privacy model
- Audio never leaves your device
- No accounts required (for DIY setups)
- No data retention policies to worry about
- No subprocessors, no GDPR transfers, no HIPAA BAAs
- You control backups, encryption, and access
Cloud AI privacy model
- Audio is uploaded to vendor servers
- Stored indefinitely unless manually deleted
- Processed by third-party infrastructure (AWS, Google Cloud)
- Subject to vendor privacy policy changes
- May be used for model training (check terms)
- Requires trust in vendor security practices
For most casual users, cloud privacy is fine. For lawyers, doctors, consultants, and anyone with confidentiality obligations, local AI isn't just preferable — it's often mandatory.
Which should you choose?
Choose local AI if:
- Privacy and confidentiality matter
- You want to eliminate subscription costs
- You work offline or on restricted networks
- You don't mind setting up software (or using a tool like Clearminutes)
- You process many hours of meetings per week
Choose cloud AI if:
- You prioritise convenience over privacy
- You need multi-device access without setup
- You want built-in team sharing and analytics
- You have light usage (under 5 hours/week)
- You don't mind ongoing subscription costs
"The local vs cloud decision isn't about technology. It's about who you trust with your data. If the answer is 'nobody but me,' local AI is the only option."
Ready to try local transcription? Download Clearminutes and get the accuracy of cloud AI with the privacy of local processing. Or read our offline transcription guide to set up Whisper.cpp yourself.