Before a team commissions code-switched speech data, someone always asks the sensible question: can we just use what is already public?
The answer depends almost entirely on which language pair you need, and the distribution is more lopsided than most people expect.
"Is there public data" is the wrong question. The useful one is "is there public data for my pair, in my register, recorded the way my users actually speak" — and those three qualifiers are where the available corpora thin out fast.
Research concentrates on three pairs
A 2025 systematic review of end-to-end code-switching ASR went through the peer-reviewed literature and recorded which languages each paper covered. Its finding on concentration is blunt: three language pairs account for roughly 76% of all papers — Mandarin–English, Hindi–English and Arabic–English.
Mandarin–English alone covers about 55%. Hindi–English is about 11%, Bengali–English about 3%.
That matters because research attention and data availability move together. The pairs that get studied are the pairs that have corpora, and the corpora get built where the research is.
For Mandarin–English, the data is genuinely there
The survey’s table of publicly available datasets shows real depth for this one pair:
| Corpus | Year | Speech | Hours |
|---|---|---|---|
| ASRU 2019 | 2020 | Read | 740 |
| TALCS | 2022 | Spontaneous | 587 |
| SEAME | 2010 | Spontaneous | 192 |
| ASCEND | 2022 | Spontaneous | 10.6 |
Hundreds of hours of spontaneous Mandarin–English is a real resource. If that is your pair, public data can carry an evaluation set and possibly more.
Two details are worth reading closely, though.
The most-used corpus is from 2010. SEAME is the single most frequently used dataset in the surveyed literature, appearing in about 24% of papers. It was published in 2010. Speech recorded in 2010 does not contain the product names, loanwords or interface vocabulary that a 2026 voice product has to recognise, and those embedded terms are exactly where code-switched recognition tends to fail.
The largest corpus is read speech. ASRU 2019 is the biggest entry in the table at 740 hours, and it is read, not spontaneous. Size and usefulness are not the same variable here.
Everywhere else, the numbers collapse
Outside those three pairs — and outside Mandarin–English within them — the picture changes completely.
Arabic is, by the survey’s count, the best-served language after Mandarin: it identifies eight Arabic code-switching datasets, more than any other language, most of them purpose-built for code-switching research. Their sizes:
| Corpus | Pair | Hours |
|---|---|---|
| TunSwitch CS | Arabic–French | 163.6 |
| Mixat | Emirati Arabic–English | 15 |
| QASR.CS | Arabic–English | 12 |
| ZAEBUC-Spoken | Arabic–English | 12 |
| ESCWA-CS | Arabic–English | 2.8 |
| HAC | Arabic–English | 2 |
Elsewhere: Frisian–Dutch (FAME) 14 hours. German 3.4 hours. Spanish 35.7 hours. A multilingual South African soap-opera corpus runs 14.3 hours in total, of which roughly 3.6 hours is actually code-switched.
Two to fifteen hours is an evaluation set. It is enough to measure a system. It is not enough to adapt one.
The broadest multilingual set contains no spontaneous code-switching
The obvious counter is CS-FLEURS, presented at Interspeech 2025, which covers 113 unique code-switched language pairs across 52 languages — far beyond anything else available. Three hundred hours.
Its composition is the part that decides whether it solves your problem:
| Subset | Pairs | How the speech was produced |
|---|---|---|
| Read-Test | 14 | Real voices reading synthetically generated sentences |
| XTTS-Train | 16 | Generative text-to-speech |
| XTTS-Test1 | 16 | Generative text-to-speech |
| XTTS-Test2 | 60 | Generative text-to-speech |
| MMS-Test | 45 | Concatenative text-to-speech |
Read or synthetic throughout. The breadth is real and the dataset is a genuine contribution to a field that badly needed coverage beyond three pairs — but not one of those 113 pairs is a recording of someone actually code-switching in conversation.
The field says this about itself
This is not a criticism from outside. The ASCEND paper opens by naming the problem directly:
Despite the spontaneous nature of code-switching in conversational spoken language, most existing works collect code-switching data from read speech instead of spontaneous speech.
The distinction matters more here than in most speech tasks. Where a speaker switches is driven by conversational context — what the other person just said, which concept has no comfortable equivalent in the base language, whether a term arrived through work or through school. A reader following a script switches where the script says to. The placement, the density and the hesitation around the switch are all different, and those are precisely the signals a model has to learn.
Public corpora can usually give you an evaluation set, and for Mandarin–English they can give you more. They rarely give you adaptation data, because adaptation data has to match your scenario, register and market variety — and a public corpus was built for someone else's.
What to do with this before you commission anything
Check your pair against the list first. If you are building for Cantonese–English, Tagalog–English, Thai–English, Vietnamese–English or Korean–English, the honest answer is that there is little or nothing public, and a vendor who claims otherwise is worth questioning closely.
Separate the evaluation question from the training question. They have different requirements and different answers. A ten-hour public corpus may be a perfectly good benchmark while being useless for adaptation. Deciding both at once is how teams end up with neither.
Match the vintage and the register, not just the language. A 2010 corpus and a 2026 product disagree about vocabulary. A read corpus and a live support line disagree about where switches fall. Neither mismatch shows up in a headline hours figure.
Know what you are buying when you buy synthetic. Generated code-switched speech is useful for augmentation and for breadth, and the research community uses it for exactly those reasons. It inherits the switch-boundary behaviour of the model that generated it, which is the behaviour you were trying to measure.
Sources
- Agro, Kulkarni, Kadaoui, Talat and Aldarmaki, Code-Switching in End-to-End Automatic Speech Recognition: A Systematic Literature Review, arXiv:2507.07741 — concentration figures and the dataset table.
- Lovenia, Cahyawijaya, Winata et al., ASCEND: A Spontaneous Chinese-English Dataset for Code-switching in Multi-turn Conversation, LREC 2022, arXiv:2112.06223.
- Yan, Hamed, Shimizu et al., CS-FLEURS: A Massively Multilingual and Code-Switched Speech Dataset, Interspeech 2025, arXiv:2509.14161.
For why this data has to be recorded rather than scraped or generated, and what a production specification contains, see code-switching speech data: why it’s the frontier multilingual AI can’t skip. Our language data work for AI teams covers code-switched speech collection and validation across Asian-language pairs, with consent and provenance documented per contributor and batch. If you are scoping a pair and want a straight answer on what exists, start here.