Brazilian Portuguese Transcription: BR vs EU Explained

The single most important thing to know about Portuguese transcription: Brazilian and European Portuguese are different enough at the phoneme level that the wrong engine setting will meaningfully hurt your output. Both share grammar and core vocabulary but diverge in pronunciation, spelling conventions, and vocabulary to a degree that breaks models trained on one variant when they encounter the other.

Brazil is the world's second-largest podcast market after the United States. More than half of Brazilians listen to podcasts, averaging 11 hours per week. Eighty-two percent prefer podcasts in Portuguese, and 81% discover new shows through YouTube. That scale creates an enormous demand for accurate Brazilian Portuguese transcription at prices that scale with output volume.

Why pt-BR and pt-PT Break Differently in ASR

The phonological gap between the two variants is large enough to fool ASR models trained on the wrong data.

Brazilian Portuguese keeps unstressed vowels fully voiced. European Portuguese reduces or drops them. The word "esperança" in Brazil sounds like "es-pe-RAN-sa" with all syllables present. In Lisbon it collapses to something closer to "shpRANsa." A model trained on Brazilian data will attempt to transcribe that reduced European Portuguese utterance as a different word entirely.

The key consonant difference for transcription is the BR palatalization of "t" and "d" before "i" or final "e". In standard Brazilian pronunciation, "bom dia" is phonetically [bõ dʒia], with the /d/ becoming an affricate. In European Portuguese the same word is [bõ dia], a clean dental stop. Models must learn to map one phoneme to the letter "d" in two very different acoustic shapes.

At the end of syllables, European Portuguese turns "s" into a hushing sound similar to English "sh." Brazilian Portuguese keeps it a clean sibilant. The word "estas" in Portugal sounds like "eshtash." The same word in Rio or São Paulo sounds like "estas." An engine expecting one will misread the other.

Vocabulary divergence compounds the problem. Common words like "bus" (ônibus in BR, autocarro in PT) and "cell phone" (celular in BR, telemóvel in PT) are not just different spellings of the same root; they share no resemblance. If a model selects the wrong variant, it may produce a correctly-spelled word from the wrong country.

The Training-Data Skew That Biases Every Engine Toward BR

Research on multilingual ASR models consistently finds a strong bias toward Brazilian Portuguese due to higher-resource pre-training data. When zero-shot models are tested on both variants, they systematically perform better on pt-BR. When researchers fine-tune on European Portuguese audio, performance on BP often degrades rather than both improving together.

This means European Portuguese users bear the accuracy penalty for an imbalance they did not cause. The practical implication: if you are transcribing European Portuguese, you need an engine that explicitly trained on pt-PT, not one that just claims "Portuguese support."

Deepgram's Nova-3 model accepts three separate language codes: pt, pt-BR, and pt-PT. As of May 2026, Deepgram documented improved accuracy across both Portuguese variants for Nova-3. That explicit variant routing matters because the model is not just applying a label after the fact; the language code shapes which acoustic patterns it listens for.

Whisper Large-v3 also supports Portuguese with a language hint, but its training-data bias skews toward pt-BR. Setting language=pt on a Lisbon recording will still run the model, but accuracy on reduced vowels and the hushing sibilants will suffer more than it would on equivalent Brazilian audio. See the Whisper architecture and language tradeoffs for a deeper look.

AssemblyAI's Universal-3 Pro model supports Portuguese, but does not offer a pt-BR or pt-PT split in the language parameter. You pass "Portuguese" and the model processes it. That is usable for Brazilian content, where the training-data advantage works in your favor, but it is a blunter instrument for European Portuguese.

The 2009 Orthographic Agreement and Why It Still Trips Up Engines

The Acordo Ortográfico came into force in 2009 with the goal of unifying Portuguese spelling across Portuguese-speaking nations. It did not fully succeed: persistent differences remain, and AI engines trained on pre-2009 or mixed-vintage text can produce spellings that are technically correct for one context but wrong for another.

Key post-2009 changes include the removal of silent consonants from European Portuguese: "acção" became "ação", "óptimo" became "ótimo". Both now match the Brazilian spelling. But stress-mark differences survive. Brazilian Portuguese uses "gênero" and "tônico" where European Portuguese spells the same words "género" and "tónico." A model generating the wrong accent placement on a word like this is not producing noise; it is producing the other variant's spelling.

For transcription, this matters most in post-processing. If you are piping output into a spellchecker or document formatter calibrated to one orthographic standard, mixed-variant output will produce false positives throughout.

What Diacritics Must Be Preserved

A correct Portuguese transcript must keep: the tilde (ã, õ), cedilla (ç), acute accent (á, é, í, ó, ú), circumflex (â, ê, ô), and grave accent (à). The most failure-prone characters are the tilde and cedilla. An engine that returns "nao" instead of "não", "voce" instead of "você", or "acao" instead of "ação" is wrong at the character level, not just stylistically imprecise. Diacritic-free Portuguese text breaks downstream systems that parse it for meaning and creates rework that negates any time savings from automation.

Whisper Large-v3 preserves diacritics reliably on clean audio. Deepgram Nova-3 also handles them correctly. Lighter, cheaper models are more likely to strip accents, especially on nasal vowels.

Code-Switching in Brazilian Tech and Business Audio

Brazilian Portuguese speakers, particularly in technology, media, and business contexts, regularly insert English words mid-sentence without switching scripts. This is not sloppy language use; it reflects how Portuguese has absorbed English vocabulary in specific domains.

Common patterns: "vou fazer um deploy antes da reunião", "precisamos dar um upgrade no servidor", "o pull request foi aprovado ontem." The Portuguese sentence structure remains intact; English nouns and verbs are inserted in their original form, often conjugated using Portuguese morphology ("deployar", "stackear").

Well-established loanwords like "internet", "site", "startup", "webinar" are fully absorbed and transcribe correctly with any good engine. The difficulty is less-common or context-specific English phrases inserted into an otherwise Portuguese sentence. Engines may output a phonetic approximation, drop the word, or produce a Portuguese word with similar sounds.

The most reliable approach is a custom vocabulary or keyterm list for recurring proper nouns, brand names, and technical terms. See the guide on fixing multilingual code-switching for a structured workflow.

Engine-by-Engine Comparison for Portuguese

Note: accuracy figures for each engine vary by audio quality, speaker accent, and domain vocabulary. The table reflects verified language-support status, not vendor-reported accuracy claims.

Pricing Snapshot (July 2026)

Tools that charge per minute become expensive fast for Brazilian creators transcribing long podcast episodes. The creator and business market in Brazil, where 81% of weekly podcast listeners use YouTube as their primary discovery platform and video podcast formats are standard, generates high-volume transcription demand.

Happy Scribe's AI tier runs from roughly $17/month for 120 minutes to $89/month for 6,000 minutes of AI credits. Trint starts at approximately $80/seat per month with a 7-file monthly cap, which limits it to occasional professional use rather than ongoing production. TurboScribe Unlimited costs $10/month on annual billing or $20/month otherwise, with no file or minute caps.

If you just need clean Portuguese transcripts without a meeting bot or editorial workflow, ConvertAudioToText offers unlimited transcription on the Pro plan at $9.99/month, with a 10-minute free tier to test the output quality on your own audio. For high-volume Brazilian podcast production, the flat-rate model consistently undercuts per-minute pricing once you exceed a few hours per month. See the transcription pricing comparison for a wider breakdown.

Workflow for Brazilian Podcast Producers

Brazil's podcast market runs on video-first production. Most Brazilian podcasters record MP4 or stream live, then distribute audio separately. The transcript serves multiple downstream uses: show notes, YouTube chapter markers, social clips, and SEO.

An efficient workflow:

1. Record the episode in your preferred format (MP4 for video podcast, MP3 for audio-only).
2. Set language to pt-BR explicitly. Auto-detect works most of the time but fails on short clips, accented speakers, and audio with background noise.
3. Enable speaker labels. Two-speaker podcasts get reliably clean attribution; three or more speakers in a casual round-table format benefit from a post-processing review pass.
4. Build a custom vocabulary for recurring terms: your show's name, guest names, sponsors, and any recurring English tech terms you use in Portuguese context.
5. Export the SRT for YouTube captions and the plain text for show notes.

The key difference from English podcast production is step 4. Brazilian tech and business podcasts have a higher rate of code-switching and proper-noun density than comparable English-language shows, and custom vocabulary pays off faster.

For interviews in European Portuguese with guests from different Portuguese-speaking regions (Angola, Mozambique, Cape Verde), the dialect picture gets more complex. African varieties of Portuguese often pattern closer to European Portuguese phonologically but with distinct vocabulary. See why AI struggles with low-resource languages for context on what to expect from those recordings.

Tips for Better Accuracy on Either Variant

1. Set the language code explicitly. pt-BR and pt-PT route to different model configurations in engines that support the split. Leaving it as auto-detect adds an inference step and occasionally mislabels short clips.
2. Record each speaker close-mic or separate. Brazilian conversational style involves more interruption and overlap than European Portuguese interview formats. Overlapping speech is the single biggest cause of attribution errors in diarization.
3. For proper nouns specific to Brazilian geography or culture, use a glossary. City names ending in "-aço" or state names like "Ceará" have common misspelling patterns. The same applies to Portuguese place names and surnames.
4. Low-reverb recording improves accuracy on both variants but matters most for European Portuguese, where unstressed vowel reduction makes phoneme boundaries harder to distinguish in echoing rooms.
5. Check for diacritic preservation before accepting any tool. Run a short test on audio containing "ação", "não", and "você". If the output strips any of these characters, switch tools before investing in a longer file.

FAQ

Why does European Portuguese transcribe worse than Brazilian Portuguese with most AI engines?

ASR models learn from whatever audio they are trained on, and Brazilian Portuguese vastly outnumbers European Portuguese in public datasets. Research confirms that zero-shot models consistently perform better on pt-BR, and fine-tuning on European Portuguese data often degrades pt-BR performance rather than lifting both evenly. Engines built explicitly for pt-PT, or the Deepgram Nova-3 model with the pt-PT language code, narrow this gap.

Does it matter which language code I set: pt, pt-BR, or pt-PT?

Yes. Deepgram Nova-3 supports all three codes and routes your audio to the appropriate variant model. Whisper Large-v3 accepts a language hint and performs better on pt-BR by default due to training-data skew. AssemblyAI's Universal model uses a single "Portuguese" option without a regional split. Setting the wrong code or using a model without variant support risks phoneme-level errors, wrong punctuation conventions, and spelling mismatches introduced by the 2009 Orthographic Agreement.

Does Otter.ai support Portuguese transcription?

No. As of mid-2026, Otter.ai supports English, Spanish, French, German, Japanese, and Simplified Chinese. Portuguese is not on that list. Brazilian creators and European Portuguese users need a different tool.

How should I handle English words mixed into a Brazilian Portuguese recording?

Brazilian speakers regularly drop English tech and business terms mid-sentence without switching scripts: "vou fazer um deploy", "temos que dar um upgrade". Most modern engines handle well-established loanwords correctly, but in-sentence mixing of less-common English phrases can produce misspellings or dropped words. Running a light review pass after transcription and building a custom glossary for recurring proper nouns or brand names is the most practical fix.

What diacritics must a Portuguese transcript preserve, and which ones do engines typically drop?

A correct Portuguese transcript must keep the tilde (ã, õ), cedilla (ç), acute accent (á, é, í, ó, ú), circumflex (â, ê, ô), and grave accent (à). The most commonly dropped characters are the tilde and the cedilla, which turn "não" into "nao" and "ação" into "acao". If your tool returns stripped output, treat it as a hard disqualifier: diacritic-free Portuguese is wrong at the character level and causes downstream errors in any system that parses the text.

Sources

• Deepgram Models and Languages Overview: https://developers.deepgram.com/docs/models-languages-overview
• Deepgram Nova-3 Portuguese expansion: https://deepgram.com/learn/deepgram-expands-nova-3-with-spanish-french-and-portuguese-support
• Deepgram changelog May 13, 2026: https://developers.deepgram.com/changelog/2026/5/13
• Otter.ai supported languages: https://help.otter.ai/hc/en-us/articles/26660468516631-Transcribe-conversations-in-English-Spanish-French-German-Japanese-or-Chinese-Simplified
• Otter.ai pricing: https://otter.ai/pricing
• ConvertAudioToText pricing: https://convertaudiototext.com/pricing
• Happy Scribe pricing: https://www.happyscribe.com/pricing
• Happy Scribe Portuguese transcription: https://www.happyscribe.com/transcribe-portuguese
• TurboScribe pricing: https://turboscribe.ai/pricing
• Brazil podcast market (Podnews, updated April 2026): https://podnews.net/article/ignoring-brazil
• Accord Ortographique 2009 spelling changes: https://www.practiceportuguese.com/learning-notes/orthographic-agreement/
• Whisper large-v3 pt-BR fine-tuned model (research reference): https://huggingface.co/freds0/distil-whisper-large-v3-ptbr
• AssemblyAI language support for streaming: https://www.assemblyai.com/docs/faq/language-support-for-real-time-transcription