How our AI works - and why it matters
We benchmarked 95 speech models across 120+ independent sources to choose the AI behind every transcript. Here is exactly how your meetings are processed - no black boxes.
95
AI speech models evaluated
120+
independent sources verified
14
languages benchmarked in depth
2
models per language: primary + fallback
The right AI model for every language
Most meeting tools use one AI model for everything. That works great for English - but accuracy drops significantly for other languages. IronMemo automatically selects the best model for each language.
Traditional approach
One model for all languages. Good for English, mediocre for everything else. Accuracy drops 40-60% for non-English content.
IronMemo approach
Language detected automatically. Best model selected for each language - with a verified fallback on standby, so your meeting is never stuck behind one provider.
3x more accurate for non-English languages
Our quality bar
The world's most popular transcription model didn't make the cut
Independent researchers and an Associated Press investigation documented that Whisper - the open-source model behind many AI notetakers - can invent entire sentences during pauses: phrases nobody said, even made-up medications in medical transcripts.
Our own silence tests confirmed the failure mode. So we banned it. Any model that writes text where there is no speech is excluded from IronMemo - no matter how popular or how cheap.
Sources: "Careless Whisper" study (ACM FAccT 2024) · AP investigation via Fortune (2024)
Silence test, mandatory
Every candidate model is fed silence, noise, and music before it sees a single real meeting. If it invents words, it is out.
Banned, not patched
We do not wrap hallucination-prone models in filters and hope for the best. They are excluded from our routing entirely.
Verified on every file
Production checks run on each transcript: no text in silent regions, timestamps in order, speakers detected. A failed check sends the file to a fallback model for reprocessing.
How a model earns its place
Most tools pick one vendor and stay. We run model selection as a continuous research program - here are the rules every candidate has to pass.
Independent benchmarks first
Vendor accuracy claims are treated as unverified until confirmed by independent benchmarks or our own tests. Every number in our research database carries a source and an evidence grade.
Meeting audio, not studio audio
We rank models on noisy, multi-speaker meeting recordings - overlapping speech, accents, bad microphones. A model that only shines on clean audiobook audio doesn't ship.
Silence and robustness tests
Candidates must transcribe silence as silence. We probe pauses, background music, and noise to screen out models that invent text - before they ever touch a customer meeting.
Every decision on record
Model choices are logged with dates, sources, and the numbers behind them. When a stronger model ships, we re-run the comparison - and can show exactly why we switched.
What happens to your data
Upload
Your recording arrives encrypted (TLS 1.3)
Process
AI transcribes and summarizes. Processing servers, no permanent storage.
Deliver
Results stored encrypted (AES-256) in your workspace
Delete
Raw audio/video deleted after processing. Only transcript and summary remain.
We never use your data to train AI models. Ever. This is not a marketing claim - it is a technical architecture decision.
Timestamps you can audit
Every line in an IronMemo transcript points to the exact second of your original recording - even on a three-hour call. Click a timestamp and the player jumps to the moment it was said. That mapping is checked on every file we process.
Word-level precision
Timestamps come from the speech model itself, not from guesswork after the fact.
Click to verify
Disagree with the summary? Jump to the source audio in one click and hear it yourself.
Holds up on long calls
Processing never silently re-cuts your audio: timecodes stay true to the original file from the first minute to the last.
Q2 planning call
Maria
Let's lock the launch date for the 21st.
Alex
Agreed - marketing needs two weeks of runway.
Maria
Final budget: 200k, signed off today.
Full transparency
AI Models We Use
Specialized speech models per language, plus LLMs from providers like OpenAI and Anthropic for summaries. Audio goes only to providers under no-training terms - and we publish what we refused to use.
Known Limitations
AI is not perfect. Transcription accuracy varies by audio quality, number of speakers, and background noise. We show confidence scores so you know when to double-check.
Human Control
You can always edit, delete, or export your data. AI assists - you decide.
Compliance and certifications
Our methodology
Our model choices run on a living research database: 95 speech models evaluated against 120+ independent sources, with evidence grades, silence tests, and a full audit trail. Champions are re-validated as new models ship. Our 3x accuracy claim for non-English languages is based on internal testing across 15 language pairs using standardized meeting recordings.
We publish our methodology and invite independent verification. Contact research@ironmemo.com for details.
Ready to stop losing what matters in your meetings?
Start for free. No credit card required.
Free plan · No credit card · Setup in 60 seconds