Portuguese Meeting Transcription Accuracy: We Tested 4 AI Tools (2026)

A solid Portuguese meeting transcription tool has to do more than just keep up with the conversation. AI meeting assistants are now a key way of taking all of the ideas and action points from your conversations in Portuguese and turning them into a way that they can fit into workflows and be easily managed, with minimal legwork.

The catch is that for Portuguese speakers, many tools are built English-first, and Brazilian Portuguese often features rapid speech, fast contractions, a sprinkle of loanwords, and is a language that an anglicized tool might struggle to keep up with.

Equally, a summary is only as good as the words that have been captured; if these are captured incorrectly in any language, it can make everything that comes after it impossible to work with.

I should say where I am coming from. I am learning Portuguese, badly, mostly through Duolingo and practicing with the team at tl;dv, and mostly at the level where I can order a coffee and lose an argument. A lot of my colleagues are Portuguese speakers and a good number of the team are Brazilian, which is how this test got scoped in the first place.

That puts me in a slightly useful position. I know enough to notice when a transcript has gone wrong, and nowhere near enough to explain why, so I could not have judged this on my own ear. It also means I have spent enough time listening to Brazilian Portuguese spoken quickly, with English product vocabulary dropped into the middle of sentences, to know that the tidy version in a language app is not the version anyone actually uses in a meeting.

To see which tools actually hold up, I put four of the best-known meeting assistants through the same test: three real Brazilian Portuguese meetings, scored blind by two LLMs and a native speaker across accuracy, real-world usefulness, features, and security. Here is how they compared, before I get into the details.

Tier	Max	tl;dv	Google Gemini	Tactiq	Read.ai
Transcription & accuracy	65	57/65	41/65	27/65	44/65
Real-world meeting quality	45	43/45	26/45	33/45	35/45
Capabilities and features	72	66/72	40/72	59/72	57/72
Trust, security and value	18	18/18	15/18	15/18	15/18
Overall score	200	184/200	122/200	134/200	151/200
Rank		1	4	3	2

Transcription Accuracy

Before I ran a single clip, I checked what each tool says about Portuguese on its own website. All four claim it, but with slightly different caveats.

Tool	What it claims about Portuguese on its own site
tl;dv	Names Portuguese for both Brazil and Portugal in its enumerated list of 30 or more languages. Platform interface localized in seven languages, Portuguese among them
Google Gemini	Portuguese is one of eight languages for the Meet saved transcript and for Take notes for me. One spoken language per meeting. No Brazilian or European variant named
Tactiq	Dedicated Portuguese transcription, speech-to-text, and translation pages. Its own site gives the language count as 30 or more, 35 or more, and 60 or more on different pages. One language per meeting, and the transcript language cannot be changed after the call
Read.ai	Portuguese is one of 26 supported languages as of March 2026. Language is auto-detected rather than set. States that meetings with more than one language spoken are not fully supported. No Brazilian or European variant named

Read that middle column, and you will see one thing stands out. tl;dv is the only tool of the four that distinguishes Brazilian Portuguese from European Portuguese. The other three list “Portuguese” as a single undifferentiated language. Those are two dialects with different vocabulary, different forms of address, and different syntax, and a tool that has not separated them in its own documentation has probably not separated them in its model either.

The second thing is code-switching. Three of the four cannot handle a meeting where speakers move between languages. Gemini supports one spoken language per meeting and says multilingual meetings are not supported. Tactiq supports one language per meeting and cannot relabel a transcript afterward. Read.ai says mixed-language meetings may produce inconsistent results. In a Brazilian business meeting, where English product and finance vocabulary drops into Portuguese sentences constantly, that limitation is not theoretical.

Portuguese meeting transcription accuracy starts and ends with getting the words spoken in any Portuguese-language meeting correct. Meetings are often busy, with ideas, figures, and action items discussed at a rapid pace, going off on tangents, and capturing all-important information needed for ongoing decisions.

Portuguese, as a language, is nuanced around diacritics, contractions like pra and tá, loan words, and acronyms. My testing confirmed this, with many of the tools struggling to capture nouns and figures when spoken rapidly, sometimes resulting in catastrophic errors.

tl;dv scored the best marks by a fairly wide margin, scored anonymously by both the LLMs (Anthropic’s Claude and OpenAI’s ChatGPT) and our native speaker. Below is the full line-by-line breakdown of how each tool performed.

Metric	How scored	tl;dv	Google Gemini	Tactiq	Read.ai
Language accuracy	Blind native-speaker severity rating on in-language accuracy	18/20	15/20	12/20	15/20
Language-specific handling	Diacritics, punctuation, regional variants, code-switching	17/20	13/20	10/20	16/20
Word error rate scoring	Computed against an official transcript or reference text	5/5	2/5	0/5	1/5
Entity detection	Names, companies and places across the cast	5/5	2/5	1/5	3/5
Numbers, dates and currency	Figures, dates and amounts formatted correctly in-language	4/5	3/5	1/5	3/5
Technical term raw recognition	Industry terms and acronyms before custom training	5/5	3/5	1/5	3/5
Punctuation and segmentation	Sentence breaks and paragraphing in test-run output	3/5	3/5	2/5	3/5
Transcription & accuracy subtotal		57/65	41/65	27/65	44/65

Six failure patterns showed up across more than one tool. Each one is a check you can run on your own transcript in under a minute, so I have written them that way rather than as a list of scores.

The Slur

The most serious single error in the test came from Tactiq, which rendered se aquele mercado é viável as se aquele mercado é ****. Viable became a racial slur. One instance, one run, in a transcript of a public policy hearing.

Our native speaker was actually sitting opposite me when she was reading through the anonymized outputs. I could see her face change before she said anything, and when I asked her to read the line back that was causing her to look upset, she would not say the word out loud. That is a native Brazilian Portuguese speaker, reading a transcript of a public policy hearing, declining to voice what the tool had put in it. It is a mishearing of one word rather than a pattern I saw repeatedly, and I have not seen it from any other tool across this series. It is still the reason I would not put Tactiq on anything high-stakes in Portuguese. A transcript that inserts a slur into a client meeting is a liability, and the tool gives you no signal it has happened.

The Diacritic Tell

Portuguese carries meaning in its accents, and they are the first thing an English-first engine drops. Gemini wrote flíguido where the speaker said fígado, the word for liver. It is a word that does not exist, sitting in a sentence about organ failure.

The layer to watch is the nasal ã, and á held apart from à. tl;dv kept them intact throughout and took 17 out of 20 on this row. Tactiq took 10.

Your check: find any accented word in your own transcript and see whether the accent survived. If the tool is substituting unaccented spellings, it is approximating Portuguese rather than processing it.

The Dropped-H Tell

The Portuguese initial h is silent, which makes it the easiest letter for a phonetic engine to lose. In the same medical clip, Gemini rendered hepatologista as epatologista. Tactiq rendered it as patologista.

Those are not the same error. Epatologista is a misspelling. Patologista is a pathologist, a different medical specialism entirely, spelled correctly, sitting in a sentence where a liver specialist was speaking. One is a typo you would catch. The other reads perfectly and is simply wrong about who was in the room.

Your check: any word starting with h. If it comes back without the h and still looks like a real word, you have the dangerous version of this error.

The Magnitude Check

Tactiq shrank a figure from R$4 bilhões to R$4 milhões. A thousandfold miss, in a formal hearing, on a monetary figure.

tl;dv’s problem on this row was different and much smaller. It writes figures as words, so 30 milhões came out as trinta milhões. Every digit correct, just slower to scan than numerals, which is why it took 4 out of 5 here rather than the full mark. Our native reviewer flagged the readability, not the accuracy.

These two errors look similar in a scorecard and are nothing alike in a meeting. One is a formatting preference. The other changes a budget by three orders of magnitude and survives into the summary, and then into whatever decision the summary informs.

Your check: every number, by hand, against the audio. There is no shortcut on this one.

The Brand-Name Tell

Tactiq turned Roche into rodovias, the word for highways. The pharmaceutical company became road infrastructure.

Sirolimus, an immunosuppressant named repeatedly in the same hearing, came back as a different phonetic guess from three of the four tools. tl;dv held both, and did it without any custom vocabulary added.

Proper nouns have been the weak point across every language I have tested, in Swedish, French, Japanese, and Spanish. They are the highest-value words in most business meetings and the ones an engine has least reason to know. A tool that guesses at them hands you a transcript that reads fluently and names the wrong company.

Your check: pick the brand, client, or product name that comes up most in your meetings and search the transcript for it. If it appears three different ways, you have found your answer.

The Invented-Word Tell

Read.ai produced mutaconografia where the speaker said ultrasonografia. That is not a mishearing of a similar word but is a completely made-up word.

Elsewhere it turned voos into bolos, then produced aeroportos que não tem bolos, airports that have no cakes, in a discussion about air routes.

When an engine loses the thread, it does not stop. It will often generate something phonetically plausible and moves on, with no flag and no gap in the text to tell you it happened.

Your check: read one page properly rather than skimming. Invented words are obvious to a native speaker and invisible to everyone else, which is exactly why they are dangerous in a transcript that gets circulated.

The English-First Root Cause

Almost every pattern above traces back to the same thing. An engine that assumes English keeps reaching for English habits when the Portuguese gets difficult: the phonetic fallback, the loanword guess, the dropped diacritic.

Read.ai dropped an English “Yeah.” into the middle of a Portuguese sentence about air routes. Small enough not to cost it a mark, and revealing about what is running underneath.

The documentation backs this up. Three of the four tools treat Portuguese as one undifferentiated language setting layered on an English-default engine, which is what the claims table above shows. tl;dv is the exception, and its Portuguese was the most legible of the four to our native reviewer, who called it the most accurate as well.

Real-World Meeting Quality

The accurate transcription of meetings was the core of what I tested, but being able to take those raw outputs and turn them into actionable outcomes matters just as much. Meetings often come with decisions, actions to be taken, and follow-ups, so in this section, I looked at how each tool handled the continuation and how easily these can be transferred into documentation that leads to progress.

Of all four tools, scored across the different areas, tl;dv achieved a score of 43, a solid lead over Read.ai, which was able to achieve 35. Tactiq followed on 33, and Gemini trailed at 26.

Metric	How scored	tl;dv	Google Gemini	Tactiq	Read.ai
Diarization quality	Correct speaker count and turn attribution vs known cast	10/10	2/10	2/10	6/10
Behavioral stability	Behavioral stability across session types	8/10	8/10	8/10	9/10
Summary quality	Usefulness of the summary and whether it stayed in the source language, with allowances for loanwords	5/5	3/5	4/5	4/5
Hallucination/insertion rate	Invented, looped or duplicated text not present in the audio. Mishearings and truncation excluded	10/10	8/10	9/10	10/10
Action item extraction	Quality of tasks and follow-ups pulled from the meeting	5/5	2/5	5/5	4/5
Auto chapters / sectioning	Does the summary break the meeting into useful sections	5/5	3/5	5/5	2/5
Real-world meeting quality subtotal		43/45	26/45	33/45	35/45

Diarization

This is how well each of the tools is able to isolate and name the different speakers from the audio. To test this, I uploaded the same clip directly to each of the platforms to see how well it was able to parse out the speech and correctly identify each speaker. tl;dv scored the full 10 points and was able to isolate and correctly attribute speech to each speaker. Read.ai, at 6, was able to correctly identify the multiple speakers, with some issues on attribution, whereas both Gemini and Tactiq allocated all speech to a single speaker and achieved a score of 2.

The clip used was a medical-based hearing, with a regulator, a doctor, and a named deputy trading turns. Where Gemini and Tactiq fused every voice into one, this means that a clinical point from the doctor can read as if the deputy said it, and indicates that the incorrect attribution could lessen the impact and authority of the notes and summaries.

The key here is that if you have more than one person speaking in a meeting, for Portuguese, Gemini and Tactiq should be relied upon with caution, based on my testing.

Hallucinations & Inserted Lines

Both Read.ai and tl;dv were able to transcribe and generate summaries and outputs that were free from hallucinations and invented lines. Tactiq and Gemini both also scored well, but slipped slightly in a few areas. Gemini produced the oddest example of the set, dropping the phrase na escola (at school) into the middle of a discussion about liver failure in gene-therapy patients.

tl;dv and Read.ai invented nothing across all three sessions, which is why they share the top mark. Inserted text is the quiet danger here, because it reads fluently and gives no signal that it was never spoken, so a reader skimming the notes has no way to spot the fiction. Read.ai had one small slip of its own, dropping an English “Yeah.” into the middle of a Portuguese sentence about air routes, minor enough not to cost it the mark.

Summaries, Action Items & Chapters

This is where I looked at the tools’ ability to take the raw transcription data and turn it into a succinct, useful, actionable output for a Portuguese speaker. tl;dv was able to take full marks across several lines, including action-item extraction, auto-chaptering, and general usefulness, giving Portuguese-language directions.

tl;dv kept all of its summaries in Portuguese rather than switching to English, which matters when the notes go back to a Brazilian team. Read.ai landed in the middle: useful summaries and action items, but weak sectioning, so its notes were harder to navigate than tl;dv’s or Tactiq’s.

Tactiq was also able to give some solid action items, isolating the correct parts into the various chapters, in spite of the issues with the core transcription. Gemini offered the weakest summary of all the tools, missing specific action items and offering the least actionable takeaways of the four.

Capabilities and Features

On top of being able to accurately transcribe in Portuguese and take those discussions and turn them into results, there are also a range of features and capabilities that can help to integrate your meeting content into day-to-day workflows.

In this section, I looked at the attributes and features that add to what each tool can do, with a top score available of 72.

tl;dv scored 66, Tactiq 59, Read.ai 57, and Gemini 40. Three of the four tools were very strong, offering a range of useful add-ons and features to make workflows run more smoothly. Gemini struggles against these specifically because it’s part of Google Workspace rather than being a standalone tool.

Metric	How scored	tl;dv	Google Gemini	Tactiq	Read.ai
Speaker naming out of the box	Auto-names real speakers on Meet, Zoom, Teams	5/5	5/5	5/5	5/5
Voice printing	Availability of voice-print training for the user’s own voice	5/5	0/5	0/5	0/5
Bot-free recording	Records via system audio without sending a bot into the call	5/5	5/5	5/5	5/5
CRM sync	Native and auto-sync	3/3	0/3	3/3	3/3
Custom notes / templates	Customizable summary formats vs a fixed output	3/3	0/3	3/3	0/3
Custom vocab / entity training	Teach industry terms and acronyms	5/5	0/5	5/5	5/5
Portuguese UI localization	Whether the product interface itself is available in Portuguese	5/5	5/5	5/5	0/5
Integrations breadth	Slack, calendar, Zapier, API	3/3	3/3	3/3	3/3
Processing speed	Time from meeting end to finished transcript	3/3	1/3	1/3	2/3
Filler-word tracking	Tracks um, eh, este without stutter-doubling, for full visibility rather than over-smoothing	3/3	0/3	0/3	3/3
Timestamp accuracy	Spot-check that timestamps land on the right moment	3/3	3/3	3/3	2/3
Translation availability	Can it translate the meeting notes, and into how many languages	3/3	3/3	3/3	3/3
Search within transcript	Search across a meeting and across the library	3/3	3/3	3/3	3/3
Transcript editing UI	Can you correct the transcript easily after the fact	3/3	3/3	3/3	3/3
Export formats	SRT, VTT, TXT, DOCX and similar	0/3	3/3	3/3	3/3
Live / real-time transcript	Is a transcript shown live during the meeting	0/3	3/3	3/3	3/3
Meeting platform coverage	Zoom, Meet, Teams, Webex coverage	3/3	0/3	3/3	3/3
Mobile app capture	Can it record in-person meetings via a mobile app	3/3	3/3	0/3	3/3
Native MCP server	Native first-party server letting AI assistants query the meeting library	5/5	0/5	5/5	5/5
Speaker label editing	Can you rename and reassign speakers after the fact	3/3	0/3	3/3	3/3
Capabilities and features subtotal		66/72	40/72	59/72	57/72

Portuguese (Brazil) UI Localization

One of the biggest challenges when working with English-first tools in languages such as Portuguese is that the transcription and outputs are often available in multiple languages, but the user interface is not. Of the tools I tested, three offer Portuguese as an interface option: tl;dv, Gemini, and Tactiq. Read.ai can still transcribe and summarize in Portuguese, but it does not offer the option to have the software in the language. This means that if you don’t speak or read another language, it can be difficult to navigate the dashboard.

Voice Printing

One feature that only tl;dv has is voice printing. This is a setting that you can opt into, where the tool learns your voice, allowing for automatic voice recognition and attribution when out-of-the-box speaker naming isn’t available.

Custom Notes & Templates

Depending on the role you have and the type of business you are in, there are different requirements and needs for your meeting output. For example, a founder and a salesperson will both hold meetings but have very different requirements and takeaways from each other. This makes customization for elements such as notes and set templates incredibly useful. A sales team can have a default meeting template set up to feed into their CRM system that picks out key items from the conversation. Of the tools I tested, only tl;dv and Tactiq have this option.

Native MCP Server

Many meetings are captured across Portuguese-speaking countries, and within them, there are many ideas, thoughts, and things to do that ultimately can end up just sitting in transcripts and meeting notes gathering dust. The ability to connect your tool to an LLM such as Claude or ChatGPT can unlock an incredible amount of opportunity to make sure that whatever the outcome of your meeting, it’s used in the most effective way possible.

tl;dv has a native MCP server that allows users to plug their meetings into their preferred LLM and draft follow-ups or question months of past calls in plain language, without exporting a single file. This is particularly useful for in-depth analysis of sales calls and forecasting, and it offers marketing and social teams the chance to mine customer insight for pain points and to confirm or inspire content, giving a quick and easy way to find out exactly what you need to know.

In addition to tl;dv, Tactiq and Read.ai both offer this as well. Despite Google leaning heavily into the MCP angle in 2026, offering MCP at the platform and developer level, it does not offer a native MCP that feeds into the popular LLMs. So, hypothetically, you can connect your Google Drive (where the transcripts and summaries live for Gemini) to a workaround, but it’s not a native MCP connection, unlike the others.

Custom Vocabulary

From my raw transcription, there were some rather amusing, but also some very serious, errors in specific words in Portuguese. This is understandable when dealing with heavy jargon or specific terminology, and in many of these instances, the ability to input custom vocabulary would be of real benefit and could have avoided some of these errors. For my testing, I did not add any custom vocabulary, but the ability to do so is on tl;dv, Read.ai, and Tactiq. Google’s Gemini does not offer this as an option.

Trust, Security

The final section of scoring for this piece was based on trust, security, and data. This is one of the most important areas of judgment around any software or tool, and is the one that is likely to get more scrutiny than others. In my assessment, I looked at areas such as regional hosting options, recognized security certifications, the position of each tool on training AI on your audio, data retention controls, the transparency of pricing, and whether there was a usable free tier.

This section is worth a total of 18 points, and the results were close, showing that each of the tools takes these areas seriously, which is reassuring for anyone who cares about how their data is protected. tl;dv was the only tool to take all 18. Read.ai, Tactiq and Gemini each finished on 15, dropping a different line in each case.

Before the scores, here is what each tool actually does with your data, stated as fact rather than a singular number.

Tool	Where meeting data is processed	Trains AI on your content	Certifications	Free tier
tl;dv	Europe. Wasabi, Google Cloud and Hetzner, all EU data centers. AI hosting selectable Europe or US	No	SOC 2 Type II, GDPR. Infrastructure certified ISO 27001, PCI DSS Level 1, SOC 1 and 2	Yes, free forever. Unlimited recording and transcription, AI notes capped at 10 meetings a month
Google Gemini	United States or Europe on supported Workspace editions. No region control on Business Starter. Brazil not offered on any edition	No. Workspace terms state content is not used for model training outside your domain without permission	Google Workspace certifications	No. Requires a paid Workspace edition, or Google AI Pro or Ultra
Tactiq	Google Cloud, no user-facing region control. Sub-processors may store in the USA	No. Azure OpenAI with content logging disabled	SOC 2 Type II, GDPR, ISO 27001, HIPAA	Yes, free forever. 10 transcripts and 5 AI credits a month
Read.ai	AWS us-east-1, Northern Virginia, USA. Other regions only via sales for large purchases	No by default. Opt-in only	SOC 2 Type 2, GDPR, Data Privacy Framework participant. HIPAA BAA on annual Enterprise+ only	Yes, 5 meeting transcripts a month

The table above is what each tool does. The one below is what that was worth on the scorecard.

 

Metric	How scored	tl;dv	Google Gemini	Tactiq	Read.ai
Data residency / regional hosting	Regional hosting options, e.g. in-region hosting on demand	3/3	3/3	0/3	0/3
Security and compliance	SOC2, ISO 27001, GDPR	3/3	3/3	3/3	3/3
AI training on user audio	Does it avoid training AI on your audio (no training scores full marks)	3/3	3/3	3/3	3/3
Data retention controls	Control over how long recordings and transcripts are kept	3/3	3/3	3/3	3/3
Price transparency	Plan prices are published rather than quote-only	3/3	3/3	3/3	3/3
Free tier / limits	Free plan availability (a free trial alone scores 0)	3/3	0/3	3/3	3/3
Trust, security and value subtotal		18/18	15/18	15/18	15/18

Which Portuguese Meeting Transcription Tools Are LGPD Compliant?

Brazil’s data protection law does not require your meeting data to stay in Brazil. Article 33 of the LGPD does something narrower: it bars transfers abroad unless one of a defined set of mechanisms covers them. So the question is not whether a tool hosts in Brazil, because none of these four does. It is which region your data lands in, and whether that region is covered.

That changed in January 2026. On the 26th, Brazil’s data protection authority adopted Resolution CD/ANPD No. 32/2026, recognizing the European Union as offering protection adequate under the LGPD. The European Commission adopted a matching decision under Article 45 of the GDPR, and both were announced together on 27 January. Personal data now moves between Brazil and the European Economic Area without additional transfer safeguards. Both decisions come up for review within four years.

There is no equivalent covering the United States. A transfer there still needs standard contractual clauses approved by the ANPD, mandatory for that route since 23 August 2025, or another Article 33 mechanism.

That puts the four tools in three different positions.

tl;dv stores in Europe, on Wasabi, Google Cloud and Hetzner data centers. Brazilian meeting data sent there sits inside the adequacy decision by default. One setting to check: tl;dv lets you choose whether the AI itself runs in Europe or the US. Leave it on Europe and the whole path stays on the adequacy route. Switch it to the US and the AI processing step sits outside it.

Google Gemini offers Europe or the United States on supported Workspace editions, so a Brazilian team can put it on the covered side. The catch is the edition. Data regions is not available on Business Starter, the plan I tested on, and Google states that users without a supported edition are not covered even if a policy is applied to them. Brazil is not an option on any edition.

Read.ai stores in AWS us-east-1, in Northern Virginia. Its help center offers an alternative location only to large buyers who contact its sales team, so for most customers the answer is the United States.

Tactiq stores in Google Cloud with no region control you can set, and its privacy policy names third parties that may hold data in the USA.

One thing worth saying plainly: no tool is LGPD compliant or non-compliant on its own. Compliance sits with you as the controller, not with the software. What a tool decides is where your data goes, and therefore which mechanism you need. tl;dv and a correctly configured Gemini put you on the adequacy route. Read.ai and Tactiq put you on the contractual clauses route, which is legal, and more paperwork. The same question in a European context is covered in our guide to GDPR compliant meeting assistants.

If your meetings touch health, biometric or other sensitive categories, Article 11 narrows the lawful bases available to you regardless of where the data sits. Worth a conversation with whoever owns privacy at your company before you point a notetaker at that kind of call.

Accreditation

Each of the tools holds some recognized security and data accreditation, such as SOC 2, ISO 27001, and GDPR compliance, so on this line, each tool scored full marks.

Free Tier Availability

When choosing a tool for Portuguese meeting transcription, it’s ideal to give it a test before you commit, especially with a language like Portuguese, where accuracy varies so much from tool to tool. tl;dv, Tactiq and Read.ai all offer a free tier, so you can run your own meetings through them and judge the Portuguese output before paying a cent. Gemini is the exception. Meeting transcription sits behind a paid Google Workspace or Gemini plan, so there is no truly free way to trial it first, and that is the reason that costs it a point here.

The Methodology

To give the fairest possible results, I ran my testing through a structured and standardized process.

The Test Set

The set covers three real Brazilian Portuguese meetings, all committee recordings from the Câmara dos Deputados, chosen to stretch the tools across different registers and subject matter.

The first was a public-sector policy session on air transport, dense with figures, percentages and sector terms, and delivered mostly as a single-speaker presentation.

The second was a public hearing on the safety of a gene therapy, the hardest of the three, with a regulator, a doctor, and a named federal deputy trading turns, plus a steady run of drug names, medical terms, and company names such as Roche and the immunosuppressant Sirolimus.

The third was a public-health debate on the taxation of tobacco, alcohol, and sugary drinks, thick with acronyms, institutions, and monetary amounts.

Each clip ran roughly ten minutes, with multiple speakers in the two hearings and a single presenter in the policy session. Accuracy and word error rate were measured against a near-verbatim Portuguese reference built from the official records and published materials, so every output sat on the same fixed ground truth.

The Review

The data that was collected was anonymized and scored against our matrix by both Claude and ChatGPT. I then asked our native speaker, Suellen Lorga, to look at the anonymized outputs to identify and rank each of the tools from a Brazilian speaker’s point of view. It was at this part of the test that I was able to confirm some of the more specific errors, reading directly against the audio.

Unlike the other tests we have completed, we were able to conduct this part in person. I do not read Portuguese well enough to judge these outputs myself, so while she was going through the extracted anonymized test I was able to view, in real time, what her reaction was (without getting involved). I was able to watch her face, and it told me quite a lot! She grimaced her way through several of them, and the grimacing turned out to be a fairly reliable guide to what the scores did afterward.

In addition to the transcripts, I also looked at the summary, and then did an additional test by uploading one of the selected clips to the tools, rather than using the live audio. This let me check the quality of the diarization when given a raw file.

The Tool Set

I tested four tools: tl;dv, Google Gemini, Tactiq, and Read.ai. Three of them, tl;dv, Tactiq, and Read.ai, are dedicated meeting tools that record a call, transcribe it, and build a searchable library around it. Google Gemini is the odd one out by design. I included it as the general-assistant option that a large number of Brazilian teams already have sitting inside Google Workspace, to see how a broad AI model copes with the same job against purpose-built rivals.

Each tool captured the three meetings live, by playing the audio as if it were a real meeting. I then ran a fourth pass on one of the clips, uploaded directly through each tool’s own upload function, to pull clean speaker separation and to give behavioral stability a second condition to be judged against.

For Gemini, which has no meeting-capture step of its own, that direct upload was a workaround, an m4a converted from the original mp4. Plans varied. tl;dv ran on Business, Gemini through a Business Starter Workspace account, and Read.ai on a paid plan. Tactiq ran on its Free plan, which is free forever rather than a trial, and caps you at 10 transcripts and 5 AI credits a month. No custom vocabulary was added on any tool.

One assumption worth killing before the results. I wondered whether the tools that handled Portuguese well would turn out to be the ones built outside the US, on the theory that a company operating in a multilingual market treats other languages as first-class. The pattern does not hold.

 

Tool	Headquarters	Region
tl;dv	Germany	Europe
Google Gemini	Mountain View, USA	US
Tactiq	Sydney, Australia	APAC
Read.ai	Seattle, USA	US

Most readers assume Tactiq is American or European. It is Australian, and its privacy policy runs on the Australian Privacy Act 1988 with GDPR applied where applicable. It finished third. Read.ai is American and finished second. The European tool won, but Google is American and beat nobody, so origin explains nothing on its own.

What separates them is whether the product was built with non-English speakers genuinely in mind. tl;dv is German-built and names Brazilian and European Portuguese separately in its own documentation. The other three treat Portuguese as one setting on an English-default engine, wherever their office happens to be.

Engine & Plan Breakdown

Each of the four tools transcribes in its own way, and how that’s achieved comes down to the engine. It can also explain a lot about the spread of the accuracy. Some run a dedicated speech engine, one uses meeting-platform captions, and one keeps its engine private. Here is what sits under each.

Tool	Underlying engine / vendor	In-house or licensed	Engine type	Plan
tl;dv	ElevenLabs	Licensed	Dedicated ASR	Business
Google Gemini	Google Gemini	In-house (Google)	LLM	Standalone Gemini app (Business Starter account)
Tactiq	Meeting-platform live captions (Google Meet / Zoom / Teams) for transcription; Microsoft Azure OpenAI (GPT) for summaries	Licensed	Caption capture + LLM	Free plan
Read.ai	Read.ai proprietary stack (underlying ASR vendor not publicly disclosed)	In-house	ASR + LLM	Pro

Scope & Caveats

I ran my testing as fairly as possible, but there will always be some elements that impact the overall scores. So here’s what to keep in mind when reading the results:

• I focused on Brazilian Portuguese, selecting clips that typified this dialect. The overall ranking is a reasonable steer for European Portuguese, too, but the exact accuracy scores would shift.
• All three meetings were formal hearings from the Câmara dos Deputados. I picked these because they’re genuinely difficult and include complex topics, jargon, and multiple speakers, but they won’t reflect every kind of meeting you might have.
• I captured the meetings live by playing the audio into each tool as if it were a real meeting, using a single-source audio feed, then ran a fourth pass by uploading one clip directly to pull clean speaker separation and to see how consistently each tool held up across both conditions.
• No custom vocabulary was added on any tool, so what you’re seeing is how each performs straight out of the box. Three tools ran on paid plans, and Tactiq ran on its Free plan.
• For the capability and trust scores, I marked each tool on whether a feature is available and works, not on how well it performs.
• These tools change all the time, so something that was missing when I tested may well have been added since.
• One note on Gemini’s residency mark. Google offers data regions on Business Standard and above, so the capability exists and scores accordingly. It is not available on Business Starter, the plan I tested on, and Google states that users without a supported edition are not covered even if a policy is applied to them.

What Is the Best Tool for Portuguese Meeting Transcription Accuracy?

On the strength of this test, tl;dv is the best Portuguese meeting transcription tool of the four I looked at, and the win isn’t by a narrow margin.

Just to reiterate that I write for tl;dv, but this has been tested in good faith, and I’ve included everything that came up. So here is where their tool lost.

tl;dv scored 0 on export formats and 0 on live transcript, and the other three took full marks on both. Read.ai beat it ever so slightly on behavioral stability, 9 to 8. Tactiq matched it outright on action item extraction and on auto-chaptering. On numbers, it dropped a mark for writing figures longhand, which our native reviewer found harder to scan.

It scored 184 out of 200, topping all four tiers. More to the point, it won on the parts that actually decide whether a transcript is accurate or not. It had the lowest word error rate, the cleanest handling of accents and technical terms, and a solid diarization score when two of its rivals blended every speaker into one. It kept its summaries in Portuguese, and it was the tool our native reviewer said she would choose during blind testing.

Read.ai is a clear second on 151. It handled accents well and stayed steady across the different meetings, so it suits a team that wants exports and a live transcript and can live with the occasional slip, like the flights that became cakes.

Tactiq came third on 134. It is a genuinely feature-rich tool, but its Portuguese recognition let it down badly, including turning a routine word into a slur, so I would be cautious using it for anything high-stakes. This is obviously a single instance, but it was serious enough to note, and not something I have come across in my testing across tl;dv and other tools.

Google Gemini finished last on 122. It is convenient if you already live in Google Workspace, but as a dedicated Portuguese transcriber, it trailed on accuracy, on speakers, and on features. Even aside from the transcription accuracy in Portuguese, its real-world functionality and lack of ability to integrate as easily as other tools would be enough to keep in mind.

If you are a Portuguese-speaking team transcribing meetings that matter, tl;dv is my pick. Try tl;dv today in your own Portuguese-language meeting and see for yourself.