Search for an automotive speech recognition dataset and the results are patents and paper PDFs, because the honest answer is uncomfortable: there are very few real in-car speech corpora, and every one at scale is Mandarin. ICMC-ASR (ICASSP 2024) and AISHELL-5 (Interspeech 2025) are modern, multi-channel, and well designed. The English and European sets, AVICAR and SPEECHDAT-CAR, date from 2000 to 2004.
This guide is the full inventory: every open and commercial in-car speech corpus we know of, with recording setups, hours, and access terms in one table, plus the cabin noise sets that partially substitute for the missing speech.
It also covers what to do about the gap. For EU-language automotive voice the working options are noise augmentation, commercial data, or commissioned collection, and each has failure modes worth knowing before you spend.
The short version: what exists and what does not
The inventory of real in-car speech corpora is short enough to fit in one table, and at scale it contains exactly two entries, both Mandarin. Everything recorded in English or a European language dates from 2000 to 2004, in vehicles and over audio channels that no longer resemble anything on the road.
That fact shapes every sourcing decision in automotive voice. If you are building for English, German, French, or any other European language, there is no downloadable car speech dataset that matches a modern cabin, so the real choice is between augmentation, commercial data, and commissioned collection. This page covers the inventory first, then the workarounds. For the wider automotive picture beyond datasets, see our automotive voice data guide.
Every real in-car speech corpus, compared
These are the corpora we know of that contain real speech recorded inside real vehicles. Simulated data and generic noisy-speech collections are excluded on purpose: the value of an in-car speech dataset is the acoustics of an actual cabin, and those do not transfer from a quiet room with noise mixed in afterwards. For open corpora outside automotive, see open speech datasets.
| Corpus | Year | Language | Hours | Recording setup | License / access |
|---|---|---|---|---|---|
| ICMC-ASR | 2024 | Mandarin | 1,000+ across channels | Multi-channel, multi-speaker recording in a real car | ICASSP 2024 challenge corpus, research access |
| AISHELL-5 | 2025 | Mandarin | 100+ | 4 far-field mics above the door handles plus near-field reference, hybrid EV, 60+ driving scenarios | Open for research |
| CI-AVSR | n/a | Cantonese | Command-scale | Audio-visual in-car command recordings | Research |
| AVICAR | ~2004 | English | n/a | Multi-microphone array in a moving car, University of Illinois | Open, aged recording conditions |
| SPEECHDAT-CAR | 2000 | Multiple EU languages | n/a | In-vehicle recordings over telephony-era channels | Legacy, effectively obsolete |
| AMIE | n/a | English | ~30 | In-car passenger utterances | Research scope |
AISHELL-5, the best-designed set
Per the AISHELL-5 paper (Interspeech 2025), the corpus was recorded through four far-field microphones mounted above the door handles plus a near-field reference channel, across more than 60 real driving scenarios. The team recorded in a hybrid EV deliberately: the moment a combustion engine cuts in or out produces a noise transition that additive augmentation reproduces badly. If you want a model of what a modern in-car corpus should look like, this is it. Note that it is open for research use, which is not the same as open for your product.
ICMC-ASR, the largest
The ICMC-ASR corpus, built for the ICASSP 2024 challenge, is the largest real in-car speech collection anywhere: multi-channel, multi-speaker recordings that add up to more than 1,000 hours of audio across the microphone channels. The challenge results that came out of it are the closest thing the field has to a public benchmark on real cabin audio. It is also Mandarin, which is the recurring problem this page keeps hitting.
The rest of the table is small, narrow, or old. CI-AVSR adds Cantonese audio-visual command recordings, useful for multimodal command work and nothing broader. AVICAR, recorded around 2004 at the University of Illinois with a multi-microphone array, is the closest thing English has, and its recording conditions predate every microphone array shipping in a car today. SPEECHDAT-CAR covered multiple EU languages in 2000, but 25-year-old vehicles and telephony-era audio make it effectively obsolete. AMIE contributes roughly 30 hours of in-car passenger utterances at research scope.
The gap: no modern English or EU-language corpus at scale
Sort the table by language and year and the pattern is hard to miss. Mandarin has two modern, multi-channel, purpose-built corpora. Cantonese has a command set. English has a corpus from around 2004 and roughly 30 research hours. The EU languages have one corpus from 2000 and nothing since.
Old corpora fail for concrete reasons, not just age. The vehicles changed: electric and hybrid drivetrains have a different noise profile from a 2000 sedan. The microphones changed: modern cars use far-field arrays with echo cancellation, not the close-talk and telephony channels of the SPEECHDAT era. The task changed: navigation, media, and assistant queries were not in anyone's prompt scripts 25 years ago. Train or evaluate on that material and you are measuring a channel that no longer exists.
So an EU-language automotive voice program has three realistic options: augment clean speech with cabin noise, buy commercial in-cabin data where it exists, or commission a collection. The rest of this guide is about choosing between them without kidding yourself.
Cabin noise datasets: the partial workaround
You cannot download modern EU-language in-car speech, but you can get the noise, and an in-cabin audio dataset that contains only noise is still useful. On the open side, MagicHub's ASR-CabNois provides 5.08 hours of cabin noise recordings for free: enough for a first augmentation pass, thin for coverage across speeds, road surfaces, and climate-control settings. On the commercial side, one vendor sells 531 hours of cabin recordings spanning 10 car models, captured with 4 microphones and 2 phones at 6 positions; access is quote-only and no pricing is published.
The gap between those two numbers matters. An augmentation recipe is only as good as the variety in its noise pool, because a model learns robustness to the noises it saw. Five hours from one setup teaches a model one cabin. Five hundred hours across ten vehicles at multiple positions starts to resemble the distribution your product will actually meet.
What noise augmentation fixes, and what it cannot
The standard recipe is simple: take clean speech in your target language, mix in cabin noise at a range of signal-to-noise ratios, and fine-tune. For additive robustness this genuinely works, and it is the cheapest lever available. Source the clean side from any corpus you hold commercial rights to; our dataset catalogue covers roughly 60 languages, licensed for commercial use and sold by the hour.
Evaluate across a sweep, not a point. Run word error rate from -5 dB to +15 dB SNR and plot the curve: published results show error rates roughly tripling for top models in car noise, and the useful information is where your model falls off, not its score at one comfortable operating point.
Then be honest about the three things additive mixing cannot produce. Far-field reverberation: speech reaching a headliner or door-pillar array has reflected off glass and trim, and that convolutional distortion is not additive noise. The Lombard effect: people speak differently in noise, louder, slower, with shifted pitch, so noise mixed into speech recorded in silence gives you quiet-room speech plus noise, not noisy speech. Device echo paths: a car plays prompts and media through its own speakers, and the echo path through a real cabin cannot be simulated by adding noise to a file.
The practical consequence: augmented data can train, but it cannot test. Your evaluation set should be real cabin audio in your deployment language, even if every training hour is synthetic. A model that has only ever been scored on augmented audio has not been measured.
Research licenses are not commercial licenses
Almost everything in the table above is research-scoped. AISHELL-5 is open for research; ICMC-ASR is distributed under challenge terms. If the model you train ships in a vehicle, a research grant does not cover you, and dataset provenance is exactly the kind of thing an automotive procurement audit asks about.
Before you build on any open set, read the actual grant. Our guide to free speech datasets you can use commercially covers which open corpora permit shipping, and speech data licensing covers the terms worth checking before you sign or download.
Augment, buy, or commission: picking a path
Augment when your base model is already strong in the target language and the gap is robustness rather than vocabulary or accent. It is fast and cheap, and it moves the additive part of the problem. Pair it with a small real-audio test set so you know what it did.
Buy when a commercial set matches your language and your cabin conditions. For Mandarin that market exists. For EU languages there is very little speech on the shelf, which is the finding of this whole page, so buying usually means buying noise, not speech.
Commission when the deployment must match: your languages, your microphone positions, your vehicle models, your driving conditions. That is the only path that produces far-field, Lombard, real-echo audio in an EU language today. Our guide to in-car speech data collection covers how such a project actually runs: vehicles, mic rigs, scenario scripts, and consent.
A sane sequence for most teams: commission a small real evaluation set first, measure the augmented model against it, then commission training data only where the SNR sweep shows augmentation ran out. That ordering spends the expensive hours exactly where the synthetic ones failed.