Portuguese is two speech recognition problems sharing one written language. European and Brazilian Portuguese differ markedly in pronunciation, and also in spelling, vocabulary, pronoun placement and number words. Because Brazil has around twenty times the population of Portugal, most public data is Brazilian, and a model that does well on Brazilian benchmarks can still struggle with a caller from Lisbon or Porto.
Key takeaways
- European Portuguese reduces and deletes unstressed vowels so heavily that telefone can sound like two syllables. Brazilian Portuguese keeps them.
- Most open Portuguese speech data is Brazilian, including the FLEURS Portuguese split. Evaluate European Portuguese separately.
- Nasal vowels and diphthongs (pão, mãe, bem) and the tilde are central to both varieties and a common source of spelling errors.
- Vocabulary and spelling differ: facto (EP) vs fato (BP) for "fact", and in Portugal fato means "suit".
- Number words differ: dezasseis (EP) vs dezesseis (BP) for 16, and bilião in Portugal is 1012 while bilhão in Brazil is 109.
What makes Portuguese hard#
Vowel reduction in European Portuguese#
European Portuguese is stress-timed: unstressed vowels shrink to a short central vowel or vanish. Telefone is pronounced roughly tlfón, pequeno as pkenu, desenvolvimento with most of its unstressed vowels gone. Consonant clusters appear in the audio that do not exist in the spelling, and word boundaries blur.
For an acoustic model trained mostly on Brazilian speech, where these vowels are pronounced clearly, European Portuguese sounds consonant-heavy and fast. Typical errors are deletions of short function words and wrong endings: desenvolvemos o projeto heard as desenvolvemos projeto, or verb endings misread because the final vowel carried the person and it was barely voiced.
Brazilian palatalisation and final consonants#
Brazilian Portuguese has its own sound changes. In most of Brazil, t and d before an i sound become affricates: tia sounds like tchia, dia like djia. Because final unstressed e is pronounced /i/, this applies to noite (noitchi) and cidade (cidadji). Syllable-final l becomes a /w/ sound, so mal (badly) and mau (bad) merge. Syllable-final r varies from a tap to a guttural [h] to silence depending on region, and is often dropped on infinitives: falar as falá.
Syllable-final s is [ʃ] in Lisbon and Rio de Janeiro and [s] in São Paulo, a distinction that crosses the national divide.
Nasal vowels and the tilde#
Portuguese has five nasal vowels and several nasal diphthongs: pão (bread), mãe (mother), põe (puts), bem (well). The tilde and the final -m mark them in writing. Verb endings depend on them. Unstressed -am and stressed -ão share the same nasal diphthong and differ only in where the stress falls, which creates past/future confusions of the kind eles falaram vs eles falarão.
Stress and written accents#
Portuguese marks irregular stress with acute, circumflex and grave accents, and they distinguish words: avó (grandmother) and avô (grandfather), está (is) and esta (this), à (to the) and a (the). In European Portuguese, the first person plural past tense takes an accent to distinguish it from the present: falámos (we spoke) versus falamos (we speak). Brazil writes both as falamos. A pipeline that strips accents before scoring will hide all of these errors.
Clitic placement#
Object pronouns go after the verb with a hyphen in Portugal and usually before it in Brazil: diz-me vs me diz (tell me), chamo-me Ana vs me chamo Ana. European Portuguese also has mesoclisis in the future and conditional, placing the pronoun inside the verb: dir-lhe-ei (I will tell him). A Brazilian-weighted model may reorder or split these forms, producing text that is intelligible but wrong for Portuguese readers.
European, Brazilian and African Portuguese#
The two standards differ at every level, not just accent:
| Feature | Portugal | Brazil |
|---|---|---|
| "bus", "train", "screen", "mobile phone" | autocarro, comboio, ecrã, telemóvel | ônibus, trem, tela, celular |
| "fact", "contact", "reception" | facto, contacto, receção | fato, contato, recepção |
| Accents on some vowels | económico, género, António | econômico, gênero, Antônio |
| Progressive aspect | estou a falar | estou falando |
| Informal "you" | tu (with 2nd-person verbs) | você (with 3rd-person verbs); tu in some regions |
The 1990 Orthographic Agreement narrowed spelling differences, removing silent consonants such as the c in acção (now ação) in Portugal, but left the differences above in place. A transcript written in Brazilian spelling for a Portuguese customer, or the reverse, will look wrong even when every word was recognised.
Portuguese is also the official language of Angola, Mozambique, Cape Verde, Guinea-Bissau, São Tomé and Príncipe and East Timor. African varieties generally follow European spelling and have accents shaped by local languages. They are very thinly represented in public speech data.
Numbers, dates and formatting#
Several number words differ between the standards, which matters for inverse text normalisation in both directions:
| Number | Portugal | Brazil |
|---|---|---|
| 16 | dezasseis | dezesseis |
| 17 | dezassete | dezessete |
| 19 | dezanove | dezenove |
| 106 | um milhão | um milhão |
| 109 | mil milhões | um bilhão |
| 1012 | um bilião | um trilhão |
The last two rows are a factor-of-a-thousand trap in financial and news transcription. Other conventions:
- Amounts: decimal comma in both. Brazil uses a point for thousands and a prefixed symbol (R$ 1.250,50); Portugal commonly uses a space and a suffixed symbol (1 250,50 €).
- Dates: day before month, 23/09/2026 or 23 de setembro de 2026. Month names are lower-case under the 1990 agreement.
- Telephone numbers: Portuguese mobile numbers have nine digits, usually read in groups of three; Brazilian numbers are commonly read with the area code and in two groups.
- The number six is often said meia (from meia dúzia, half a dozen) in Brazil when reading digits, to avoid confusion with três. An ITN rule unaware of this writes "meia" instead of 6.
Data availability and public test sets#
Portuguese has a large speaker population and a reasonable amount of open data, but it is heavily skewed to Brazil.
- FLEURS: the Portuguese split is labelled Brazilian Portuguese (
pt_br). There is no European Portuguese split. - Common Voice: a crowd-sourced Portuguese subset, mostly Brazilian speakers.
- Multilingual LibriSpeech: a modest Portuguese portion of audiobook data.
- VoxPopuli: European Parliament speech, which is one of the few open sources dominated by European Portuguese.
- CORAA: a large, manually validated corpus of Brazilian Portuguese, including spontaneous speech.
The practical consequence: a Portuguese benchmark figure is almost always a Brazilian figure. If you serve Portugal, build your own test set from European Portuguese audio before choosing a system.
How Reson8 approaches this#
Portuguese is one of the nine European languages Resonant-1 supports in depth. Reson8 has not published a FLEURS figure for Portuguese, and FLEURS would measure Brazilian Portuguese in any case. Test on audio from the variety you serve.
Common questions#
Is one model enough for Portugal and Brazil?
For recognition, often yes, if it has seen enough of both. Output spelling, vocabulary and number formatting should still be set per locale.
Why is European Portuguese harder than Brazilian?
Mostly vowel reduction, which removes acoustic material, and less training data. Neither variety is intrinsically harder; data volume drives much of the gap.
Can a Portuguese model handle Spanish speakers, or the reverse?
No. The languages are close in writing but differ considerably in sound, and a model will produce plausible words from the wrong language.
References#
- Conneau, A. et al. (2022). FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech. arXiv:2205.12446.
- Ardila, R. et al. (2020). Common Voice: A Massively-Multilingual Speech Corpus. LREC 2020.
- Pratap, V. et al. (2020). MLS: A Large-Scale Multilingual Dataset for Speech Research. Interspeech 2020.
- Wang, C. et al. (2021). VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation. ACL 2021.
- Candido Junior, A. et al. (2023). CORAA ASR: a large corpus of spontaneous and prepared speech manually validated for speech recognition in Brazilian Portuguese. Language Resources and Evaluation.
- Mateus, M. H., & d'Andrade, E. (2000). The Phonology of Portuguese. Oxford University Press.