Transcription in 30+ languages, not one model for all of them
IronMemo transcribes meetings and recordings in 30+ languages. Instead of running one universal engine on everything, we scored 95 speech models against 120+ independent sources and route each language to the model that performed best for it, with a second model on standby. English is the easy part - the difference shows up everywhere else.
Updated July 25, 2026
- 30+ languages
- 95 models benchmarked
- Files up to 2 GB
Languages we transcribe
Each card names what makes that language hard for speech recognition - which is precisely why one universal model cannot be the best at all of them.
English
EnglishThe best-covered language in every engine. What still breaks it: crosstalk, strong regional accents and domain jargon.
Spanish
EspañolFast delivery and wide regional variation - Rioplatense, Caribbean and Castilian differ in rhythm, vocabulary and pronoun use.
German
DeutschCompound nouns written without spaces, and verbs parked at the end of long subordinate clauses.
French
FrançaisLiaison and elision erase word boundaries - connected speech runs several words together into one sound.
Portuguese
PortuguêsBrazilian and European variants diverge sharply, and European Portuguese swallows unstressed vowels almost entirely.
Japanese
日本語No spaces between words, three scripts mixed in one sentence, and pitch accent that changes meaning.
Chinese (Mandarin)
中文Tone carries meaning, and written Chinese offers no word boundaries to fall back on.
Korean
한국어Agglutinative endings plus speech levels - the same verb changes shape depending on who is speaking to whom.
Italian
ItalianoVowel-final words and fast, overlapping delivery blur where one word ends and the next begins.
Dutch
NederlandsGuttural consonants, and business meetings that switch into English mid-sentence out of pure habit.
Polish
PolskiDense consonant clusters and seven cases, with palatalization shifting sounds inside the word.
Turkish
TürkçeAgglutination: a single Turkish word can carry what English needs an entire clause to say.
Hindi
हिन्दीConstant English code-switching - Hinglish is ordinary business speech here, not an edge case.
Arabic
العربيةModern Standard Arabic and the spoken dialects differ more than some European languages do, and short vowels are not written at all.
IronMemo transcribes 30+ languages. The fifteen named above - English, Spanish, Russian, German, French, Portuguese, Japanese, Chinese (Mandarin), Korean, Italian, Dutch, Polish, Turkish, Hindi and Arabic - cover the bulk of business meetings, and the remaining languages run through the same pipeline. Updated July 2026.
When one meeting has two languages
Most notetakers make you pick a language up front. Here is what that means in practice, in their own words.
What most tools do
- Read.ai states plainly that meetings where several languages are spoken are not fully supported and may produce inconsistent results.
- Circleback detects the predominantly spoken language and transcribes the entire meeting in it - the second language is processed as if it were the first.
- Krisp asks you to select one language up front, and the transcript is generated in that language whatever is actually said.
- Fireflies does detect switches at word level, but its multi-language mode is limited to the Business and Enterprise plans.
What IronMemo does
- Language is detected as the conversation goes, so a switch partway through does not restart or corrupt the transcript.
- Russian-English and Spanish-English switching inside a single sentence is a case we tuned for, not one we merely tolerate.
- Speaker labels survive the switch - diarization does not reset when the language changes.
- Every line keeps its timestamp, so you can click back into the audio and check any sentence yourself.
Fair warning: mid-sentence switching is genuinely hard for every engine, and Sonix says so publicly - their detection works best when languages change between speakers rather than mid-sentence. If your meetings run in one widely supported language, most tools listed here will serve you well; the difference appears when they do not. Vendor statements checked in July 2026.
How a language gets its model
The short version. The full methodology, including which models we banned and why, lives on our AI page.
Benchmark first, marketing never
95 speech models scored against 120+ independent sources - accuracy, behavior on silence, diarization quality and cost, measured per language rather than in aggregate.
Route each language to its winner
Fourteen languages are pinned to a primary model with a second on standby. The rest run on the strongest general-purpose model from the same benchmark.
Keep the result auditable
Every transcript line stays clickable back to its own second of audio, so an accuracy claim is something you verify yourself rather than take on trust.
Counting languages is the easy part
Language totals exactly as each vendor advertises them - next to what actually happens to an individual language.
| Service | Languages advertised |
|---|---|
| IronMemo | 30+ |
| HappyScribe | 150+ |
| Fireflies | 100+ |
| Notta | 58 |
| Sonix | 54+ |
Totals are what each vendor advertised on its own pages in July 2026: happyscribe.com/languages, the Fireflies help center, notta.ai pricing and sonix.ai/languages. A larger number is not a worse product - if your meetings run in one widely supported language, several of these will do the job. The count stops being the answer when your language is the one handed to a generic model.
What the benchmark actually is
The numbers behind per-language routing - a research database, not a slogan.
95
speech models evaluated
Scored on accuracy, behavior on silence, diarization and cost before any of them ever touched a customer recording.
120+
independent sources
Academic benchmarks, vendor documentation and our own test corpus, with an evidence grade attached to every result.
14
languages benchmarked in depth
Each with a primary model and a fallback. The remaining languages run on the strongest general model in the same benchmark.
Questions about language support
Explore more
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