Spanish is one of the easier major languages for speech recognition: its spelling is close to phonetic, its vowel system is small and stable, and there is a great deal of training data. The difficulty lies in variety. Spanish is the official language of some twenty countries, and a model has to cope with differences in pronunciation, vocabulary, verb forms and number formatting between Madrid, Mexico City, Buenos Aires and Bogotá.
Key takeaways
- Spelling follows sound closely, so most errors come from a small set of spelling ambiguities: b/v, silent h, ll/y and written accents.
- Written accents carry meaning (esta vs está, si vs sí) but are often stripped by WER normalisers, hiding real errors.
- Peninsular and Latin American Spanish differ in pronunciation, pronouns (vosotros, vos) and vocabulary. Test on the variety you serve.
- Number formatting is not uniform: Spain uses a decimal comma, Mexico a decimal point. Un billón is 1012, not 109.
- Reson8's published FLEURS Spanish WER is 2.67%, its lowest figure across languages.
What makes Spanish hard (and what does not)#
Spanish has five vowels, each pronounced roughly the same whatever the stress, and a mostly one-to-one mapping between letters and sounds. Compared with English or French, far fewer errors arise from sounds that the spelling hides. The problems that remain are concentrated in a few places.
Spelling ambiguities#
- b and v are the same sound in every modern variety. tuvo (had) and tubo (tube), or votar (to vote) and botar (to throw away), are resolved only by context.
- Silent h: hola and ola (hello, wave), hecho and echo (done, I throw), haya and halla and, with yeísmo, aya.
- ll and y: most speakers have merged them, so calló (fell silent) and cayó (fell) sound the same.
- c, z and s: in northern and central Spain casa (house) and caza (hunt) differ; in Latin America and much of Andalusia they do not.
Written accents#
The acute accent marks irregular stress and distinguishes pairs of words: tu/tú, el/él, si/sí, mas/más, esta/está, and question words such as que/qué and donde/dónde. Stress differences are audible in some pairs (esta vs está, papa vs papá) and inaudible in others (el vs él). A transcript that reads el dijo que si for él dijo que sí has two meaning-bearing errors. Many scoring pipelines strip diacritics before computing WER, which makes such a system look better than it is. If accents matter to your readers, score with them.
Consonant weakening#
In Andalusia, the Canary Islands and much of the Caribbean, syllable-final s is aspirated or dropped: los amigos is heard as something close to loh amigo. Because the plural in Spanish is marked mainly by that s, the model has to recover number from context, a problem similar to the silent endings of French. Intervocalic d weakens or disappears in casual speech everywhere (cansado as cansao), and some Caribbean varieties interchange final r and l.
Speaking rate and clitics#
Spanish is spoken at a high syllable rate, and object pronouns attach to infinitives and imperatives as one written word: dámelo (give it to me), enviárselo (to send it to him). Splitting these produces da me lo, three words for one, and a WER penalty larger than the error deserves.
Peninsular and Latin American Spanish#
Spain has fewer than a tenth of the world's native Spanish speakers; Mexico alone has more than twice as many. Most large Spanish corpora are therefore Latin American-weighted, while European deployments serve Peninsular speakers. The main differences:
| Feature | Spain (central/north) | Latin America (general) | Rioplatense |
|---|---|---|---|
| c/z before e, i | /θ/ (distinción) | /s/ (seseo) | /s/ |
| Informal plural "you" | vosotros tenéis | ustedes tienen | ustedes tienen |
| Informal singular "you" | tú tienes | tú tienes (vos in parts of Central America) | vos tenés |
| ll / y | merged /ʝ/ | merged /ʝ/ | /ʃ/ or /ʒ/ |
| "computer", "car", "mobile phone" | ordenador, coche, móvil | computadora, carro or auto, celular | computadora, auto, celular |
A model that has seen mostly Latin American data may transcribe Peninsular vosotros forms less reliably, and vice versa for voseo. Vocabulary is less of a risk than it seems for recognition, since the model only has to write what it hears, but it matters for downstream LLM processing and search.
Co-official languages in Spain#
Catalan, Galician and Basque are co-official in their regions. Speakers switch into Spanish mid-sentence and borrow names, street addresses and institutional terms. Catalan in particular is close enough to Spanish that a Spanish model will produce plausible but wrong Spanish words for Catalan speech. See code-switching.
Numbers, dates and formatting#
Spanish number words are regular, but inverse text normalisation must know the target country.
- Decimal separator: Spain, Argentina, Colombia and Chile use a comma (3,5); Mexico and much of Central America use a point (3.5). The RAE's 2010 Ortografía accepts both and recommends a space rather than a point or comma for thousands (1 250,50).
- Large numbers: mil millones is 109; un billón is 1012. Translating billón as "billion" is a factor-of-a-thousand error that appears in financial transcripts and news.
- Currency: Spain writes 1250,50 € with the symbol after; many Latin American currencies use a prefixed $, which means pesos, not dollars.
- Dates: day before month, 23/09/2026 or 23 de septiembre de 2026. Month names are lower-case.
- Ordinals: 1.º, 2.ª with gender marking.
Data availability and public test sets#
Spanish is among the best-resourced languages after English. Public sets include:
- FLEURS (Spanish split): read Wikipedia sentences, with speakers of mixed origin.
- Common Voice: a large Spanish subset with self-reported accent labels, useful for slicing results by region.
- Multilingual LibriSpeech: several hundred hours of Spanish audiobooks.
- VoxPopuli: European Parliament speech, mainly Peninsular.
Benchmark figures for Spanish tend to be lower than for French or German on the same test sets. That reflects the transparent spelling and the data volume, not necessarily easier production audio. Accented and telephone Spanish remains harder, and regional slices of your own data will say more than a single average. See building your own ASR test set.
How Reson8 approaches this#
Spanish is one of the nine languages Resonant-1 supports. On FLEURS Spanish it scores 2.67% WER, its lowest figure across languages and the lowest in the comparison in Reson8's benchmarks post. That is read speech; conversational and regional audio will score higher.
Common questions#
Is Spanish easier for ASR than English?
On spelling, yes: there are fewer ways to write a given sound. On accents, the range is comparable. Benchmark numbers for Spanish are often lower, but compare on matched test sets only.
Do I need separate models for Spain and Latin America?
Usually not for recognition. You do need locale-specific formatting (decimal separators, currency) and an evaluation set that covers each region you serve.
Why are accents missing from my Spanish transcripts?
Some systems or post-processing steps strip diacritics. If they are present in the model output, check your normaliser and display code before blaming the model.
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.
- Real Academia Española & Asociación de Academias de la Lengua Española (2010). Ortografía de la lengua española. Madrid: Espasa.
- Hualde, J. I. (2005). The Sounds of Spanish. Cambridge University Press.