Multilingual meeting notes that actually make sense

A product team in Tokyo finishes a 40-minute meeting with engineers in Mexico City. The meeting happened in a mix of Japanese, Spanish, and English. Somebody needs to write up the notes.

What actually happens: the most fluent English speaker writes a summary in English. The Japanese context around a specific UX concern gets reduced to one bullet point. The nuance of the Spanish-language discussion about API rate limits disappears entirely. Three days later, someone in Mexico City searches for "límites de velocidad" in the meeting archive and finds nothing, because the notes only exist in English.

This is the default state of multilingual meeting documentation. We built MangoFinch to fix it.

The dominant-language problem

In any multilingual meeting, notes gravitate toward one language. Usually English. Everything said in other languages gets filtered through one person's comprehension, which means it gets summarized, paraphrased, and frequently distorted.

I have watched this happen repeatedly during our beta testing. A Japanese participant makes a precise technical point about database indexing strategy, using specific terminology. The English note-taker writes "Tanaka mentioned some concerns about database performance." The precision is gone. The specific indexing strategy — the part that actually mattered — did not survive the language crossing.

How MangoFinch preserves both sides

Every utterance in a MangoFinch transcript exists in two forms: the original spoken language and the translated version. They sit side by side.

When Kenji says something in Japanese, the transcript shows his original Japanese text with the English (or Spanish, or Portuguese) translation directly below it. When Maria responds in Spanish, her Spanish text appears with translations underneath. The conversation flow stays intact.

This dual representation matters for three reasons.

Terminology preservation. Technical terms and domain-specific language often do not translate cleanly. When a Japanese engineer says "負荷テスト" (load testing), the translation shows "load testing" — but the original is right there for anyone who needs to verify the exact term used.

Tone and intent. Translation captures meaning but often flattens tone. A polite Japanese disagreement and a direct German objection might both translate to "I don't think that will work" in English. Preserving the original lets bilingual participants gauge the actual sentiment.

Accountability. When meeting notes become action items, people need to know exactly what was said. The original language text is the ground truth.

Searchable across every language

MangoFinch indexes both the original text and every translation. When you search for a term, you get results regardless of which language it was spoken in or which language you are searching in.

During our beta, a Japanese team lead searched for "sprint planning" in English and found a segment from a meeting where the topic was discussed entirely in Spanish. The Spanish discussion about "planificación del sprint" showed up because the English translation was indexed alongside the original.

We index roughly 840 tokens per minute of meeting audio across all language variants. A typical 30-minute meeting generates about 25,000 searchable tokens.

Transcript versus notes

A transcript is everything that was said. Notes are the parts that matter.

Most meeting tools give you a transcript and call it done. But a 45-minute meeting transcript is 6,000-8,000 words. Nobody reads that.

MangoFinch generates both. The full transcript lives in the archive. On top of that, we extract structured notes with three sections:

Key decisions. Statements where participants agreed on a direction. In our beta data, meetings average 3.7 identifiable decisions per 30 minutes.

Action items. Tasks assigned to specific people with deadlines when mentioned. Each action item links back to the exact transcript moment where it was assigned — in the original language and translation.

Open questions. Topics raised but not resolved. In 68% of our beta meetings, at least one open question from a previous meeting was re-raised later.

Every element in the structured notes is available in every language the meeting touched.

How teams actually use the output

We have been watching usage patterns across 23 beta teams for 11 weeks.

The Monday review. Several teams review the previous week's multilingual meetings. Managers scan structured notes in their native language, then dive into specific transcript segments when something needs clarification. Average time: 12 minutes per meeting reviewed, compared to 25-30 minutes with raw transcripts from other tools.

Cross-language action item tracking. A consulting firm with teams in Tokyo, London, and Buenos Aires uses MangoFinch's action items as their single source of truth. Before, action items from Japanese meetings were manually translated into English for the project tracker — a 2-3 day delay. Now the English action items are available within minutes.

The "what did they actually say" lookup. This happens at least once per week per team. Someone reads a translated note, wants to verify the original phrasing, and clicks through to the transcript.

Export formats

Teams need meeting data in their existing tools. MangoFinch exports in four formats:

Markdown — full transcript with original and translated text in collapsible sections. Works in Notion, Obsidian, GitHub.

JSON — structured data with timestamps, speaker IDs, language codes, original text, translations, and action items.

Plain text — single-language simplified view.

CSV — one row per utterance with timestamp, speaker, original language, original text, translated text.

The most requested format we have not built yet: direct Jira ticket creation from action items. It is on the roadmap. Three beta teams have already automated that connection using the JSON export.

What does not work yet

MangoFinch's note extraction works best with structured meetings. Highly informal conversations and brainstorming sessions with lots of cross-talk produce noisier results.

Our action item detection has an 81% precision rate. Roughly 1 in 5 detected action items is either a false positive or has the wrong owner assigned. We display confidence scores alongside each item.

Speaker identification in mixed-language meetings is still imperfect. When two people speak the same language with similar vocal characteristics, the system sometimes attributes text to the wrong speaker. We are improving this with enrollment-based speaker profiles.

What gets used most

I track which features drive repeat usage. The top three by weekly active usage:

Cross-language search — used by 89% of active teams weekly.

Action item export — used by 74% of active teams weekly.

Original-language lookup from translated notes — used by 67% of active teams weekly.

The full transcript view ranks fifth. People want the structured output. The complete transcript serves as a reference layer, not the primary interface. That insight shaped our entire product direction.