"Spanish" and "Portuguese" are not one target each. Both languages have major national varieties that differ in pronunciation, vocabulary and grammar, and the differences are large enough to cost accuracy when a speech model hears a variety it was not trained on.
The rule is simple: match the training and test data to where your users live. That means European Spanish for Spain, Latin American Spanish for the Americas, Brazilian Portuguese for Brazil and European Portuguese for Portugal. If you serve two markets, buy both varieties and score each one separately.
This guide covers what actually differs, which variety fits which market (including the markets our catalog does not cover), what to change in the transcription spec, and how to test. Each fact about the languages links to the primary source it comes from.
Spanish: Spain and Latin America
The Real Academia Española and the other Spanish-language academies document the regional differences in their Diccionario panhispánico de dudas and their dictionary. Five matter most for speech data.
Distinción and seseo. In the speech of central, northern and eastern Spain, z, and c before e or i, have a sound of their own, the [θ] of English "think", separate from s. Seseo, pronouncing those letters as s, is general throughout Spanish America and in the Canary Islands and part of Andalusia, and the academies fully accept it. For a seseo speaker, caza (hunt) and casa (house) sound the same, so the spelling has to come from context.
Vosotros and ustedes. Most of Spain uses vosotros as the informal plural "you", with verb forms such as calláis and escribisteis. Those forms are not used in the Americas, the Canary Islands or western Andalusia, where ustedes covers formal and informal plural address alike.
Vos. In Argentina, Uruguay and Paraguay, vos replaces tú across all social classes, with verb forms like vos llegás. It is also general in Guatemala, El Salvador, Honduras, Nicaragua and Costa Rica. Mexico, by contrast, uses tú, apart from some speakers in Tabasco and Chiapas.
Aspirated and dropped s. In southern Spain and in large parts of Spanish America, an s at the end of a syllable is often aspirated, as in [pehkádo] for pescado, and in many areas a final s disappears: los hombres can come out as [lo ómbre]. Final s marks plurals and the tú form of verbs, so a model has to recover it from context.
Vocabulary. The academies' dictionary labels ordenador as the Spain word for a computer, where Latin Americans say computadora. Celular for a mobile phone is mainly American usage, and carro means a car in the Americas, where Spain says coche. Each of these shifts the word probabilities a recognizer has learned.
Mexican Spanish is one Latin American variety, not all of them
Mexican Spanish shares what separates Latin America from Spain: seseo, ustedes and American vocabulary. It does not teach a model the vos verb forms of the Southern Cone and Central America, and accents and everyday vocabulary vary from country to country. Treat it as the right data for Mexico and a sound base for the region, then add speakers from your other markets, at least in the test set.
Spanish in the United States
The US Census Bureau estimates 70.1 million Hispanic residents as of 1 July 2025, 21% of the population. Not all of them speak Spanish, but their origins tell you which Spanish to plan for. Pew Research Center estimated that 60% of US Hispanics were of Mexican origin in 2021, and it profiles 16 other origin groups, from Puerto Ricans and Cubans to Salvadorans and Venezuelans. Mexican Spanish is the natural base for a US product, with Caribbean and Central American speakers added for the rest.
Portuguese: Brazil, Portugal and Angola
Vowels. European Portuguese reduces its unstressed vowels: [e] and [ɛ] become [ɨ], [o] and [ɔ] become [u], and [a] is raised to [ɐ]. Unlike in Brazilian Portuguese, the pattern is the same before and after the stressed syllable, as a 2023 study in the Journal of the International Phonetic Association summarizes. That is much of why European Portuguese sounds clipped to anyone used to Brazilian speech.
Pronouns. In Portugal, tu and você split according to the relationship between the speakers. In Brazil, você is the form across a large central area, and other regions use tu and você side by side as variants (Nascimento, Duarte and Mendes, 2018). That changes which verb forms turn up in casual conversation.
Vocabulary. The Priberam dictionary marks autocarro (bus) and telemóvel (mobile phone) as Portugal words, with ônibus and celular as the Brazilian equivalents. For a train, Portugal says comboio and Brazil says trem.
The research gap. The INESC-ID team behind the CAMÕES benchmark (ASRU 2025) found that Portuguese speech recognition resources are mostly focused on Brazilian Portuguese. Their best models, trained or fine-tuned on 425 hours of European Portuguese, improved word error rate by more than 35% relative to the strongest zero-shot foundation model on a 46-hour European Portuguese test set. It is a measure of what variety-matched data can be worth.
Angolan Portuguese
In Angola's 2014 census, 71.1% of the population habitually spoke Portuguese at home: 85% in urban areas and 49% in rural ones. Umbundu (23%), Kikongo (8.2%) and Kimbundu (7.8%) followed, and the shares add up to well over 100%, so many people speak more than one language at home. Vocabulary differs too: Priberam lists candongueiro as the Angolan word for a privately owned shared passenger vehicle, while in informal Portuguese from Portugal it means a ticket tout.
The CAMÕES authors describe African and Asian varieties of Portuguese as seldom considered in speech research. Treat Angolan Portuguese as its own target, and test it separately rather than assuming European or Brazilian data covers it.
The five datasets side by side
SpeechData.ai has two Spanish and three Portuguese datasets. Hours, speakers and prices come from each dataset page; the last column is hours times the list price.
| Dataset | Recorded in | Hours | Speakers | Price per hour | Full dataset at list price |
|---|---|---|---|---|---|
| Spanish (European) | Spain | 1,000 | 100 | $85 | $85,000 |
| Spanish (Latin American) | Mexico | 2,000 | 200 | $65 | $130,000 |
| Portuguese (European) | Portugal | 1,000 | 100 | $85 | $85,000 |
| Portuguese (Brazilian) | Brazil | 2,000 | 200 | $65 | $130,000 |
| Portuguese (African) | Angola | 500 | 50 | $65 | $32,500 |
Together that is 3,000 hours of Spanish and 3,500 hours of Portuguese from 650 speakers. All five are spontaneous two-person conversations recorded in-country, with one channel per speaker, time-aligned and human-verified transcripts, and age, gender, region and recording device metadata for every speaker. Every speaker signed explicit commercial-use consent, and the full consent chain is provided for legal review.
Note that the Latin American Spanish dataset is recorded in Mexico, so it is Mexican Spanish. Every speaker's region is in the metadata, so ask about the regional mix when you request a free sample.
A worked example: a voice agent for Spain and Mexico with 500 hours of each variety costs 500 × $85 plus 500 × $65, which is $75,000 at list price. Volume and multi-language discounts are available on request.
Which variety to buy for each market
| Users in | Train and test on | In our catalog | Notes |
|---|---|---|---|
| Spain | European Spanish | Spanish (European), recorded in Spain | Check the regional mix: parts of Andalusia and the Canary Islands use seseo and ustedes |
| Mexico | Mexican Spanish | Spanish (Latin American), recorded in Mexico | Direct match |
| Argentina, Uruguay, Paraguay | Río de la Plata Spanish | Not covered | Vos and its verb forms in speech and writing; record locally |
| Colombia, Chile, Peru, Central America, Caribbean | The local variety | Not covered | Mexican data as a regional base, plus local speakers in training and test |
| United States | US Spanish, mostly of Mexican origin | Spanish (Latin American) | Add Caribbean and Central American speakers; plan for English code-switching |
| Brazil | Brazilian Portuguese | Portuguese (Brazilian), recorded in Brazil | Check the regional mix in the speaker metadata |
| Portugal | European Portuguese | Portuguese (European), recorded in Portugal | Brazilian data is not a substitute |
| Angola | Angolan Portuguese | Portuguese (African), recorded in Angola | Direct match |
| Mozambique and other Lusophone African countries | The local variety | Not covered | Angolan data is a different country's variety; record locally |
Where the catalog has no match, such as Argentina, Colombia or Mozambique, native speakers can be recorded to your spec through custom collection, with recruitment feasibility confirmed before the quote.
What changes in the transcription spec
A shared language does not mean a shared transcription guideline. Add these lines to the spec for each variety:
- Name the variety and the spelling norm. The 1990 Orthographic Agreement narrowed the gap between Portugal and Brazil, but its own text keeps double spellings where educated pronunciation differs: facto and receção in Portugal, fato and recepção in Brazil, and accents such as género and gênero or António and Antônio. Brazil's transition period ended on 31 December 2015. Angola is a signatory, but state the norm you want for Angolan transcripts anyway.
- Write standard spelling over reduced sounds. When a speaker aspirates or drops a final s, the transcript should still read los hombres. Decide whether annotators also tag the pronunciation, and say so in the guideline.
- Fix how numbers are written. If numbers go in as digits, the decimal separator differs by country: Spain, Argentina, Chile and Colombia use a comma, while Mexico, Puerto Rico, Venezuela and Spanish speakers in the United States use a point. Large numbers differ too: billón is a million million, but the academies' dictionary records a US sense of a thousand million. In Portuguese, Priberam defines bilião as a million million and marks the thousand-million meaning as Brazilian.
- Keep each variety's words and verb forms. Annotators should never "correct" vosotros forms, vos forms, telemóvel or ordenador into another variety. Use annotators from the same country as the speakers.
- Decide how English is written. US Spanish speakers switch between Spanish and English, and English loanwords turn up in casual speech in other markets too. Specify whether English words keep English spelling and whether they are tagged, as explained in our guide to code-switching in speech recognition.
Test each variety on its own
Report word error rate per variety, never only pooled. A pooled number is dominated by whichever variety has the most test audio, so a model can look fine overall while failing the smaller market.
Differences show up even inside one country. A 2026 study of Whisper large-v3 on recordings of elderly rural speakers found about 20% more errors for speakers in Andalusia and 24% more in Extremadura than in the Castilian heartland (incidence rate ratios of 1.20 and 1.24).
Build a held-out test set for each variety you serve, from speakers who are not in the training data, recorded on the devices and channels your product will hear, and set a pass mark for each one. Our guide to ASR test sets covers how to size and build them.
All five datasets, along with the rest of our 60 languages and dialects, are in the SpeechData.ai catalog, where you can request a free sample of each variety.
