Automotive speech datasets differ in language, task, recording setup and access terms. ICMC-ASR and AISHELL-5 document Mandarin in-car speech; AMIE describes English passenger interactions. Older resources such as AVICAR and SPEECHDAT-CAR also deserve a task-specific assessment rather than rejection based on age alone.
This is a representative comparison, not a complete market inventory. Check the actual release, duration basis, language coverage and permission terms before treating a corpus as a fit for training or evaluation.
Representative in-car speech resources
Start with the task you need to test: single-speaker commands, overlapping conversation, passenger intent or seat attribution. A corpus useful for one task may lack the labels or recording conditions needed for another. The resources below are examples to investigate, not a ranking of suppliers or a claim that other datasets do not exist.
| Resource | Published description | Reported scope | Access check |
|---|---|---|---|
| ICMC-ASR | 2024 challenge paper | Over 100 hours of multi-channel Mandarin speech and 40 hours of noise in the paper; the challenge website advertises larger totals | Check the challenge data user agreement and the release supplied |
| AISHELL-5 | 2025 paper and OpenSLR release | Over 100 hours of Mandarin speech, four far-field channels and separate noise recordings | OpenSLR lists CC BY-SA 4.0 and downloadable splits |
| AMIE | 2020 paper | 30 hours of multimodal in-cabin passenger interaction data | The paper describes the study; confirm data access and reuse terms with the authors |
| AVICAR | 2004 paper | English in-car speech recorded with multiple microphones and cameras | Confirm current access and license with the maintainer |
| SpeechDat-Car | Original project paper | Multilingual in-car recordings, with microphone and GSM channels | Check the language-specific release and its terms |
Read duration and release claims carefully
The ICMC-ASR paper reports over 100 hours of speech. Its segmented near-field-duration table lists 94.75 hours for training, 3.33 for development and 3.30 and 3.58 for the evaluation sets. The challenge website advertises 1,000 hours. These descriptions should not be treated as interchangeable totals. Ask which release and duration calculation a proposed comparison uses; do not infer unique conversational hours by multiplying microphone channels.
The AISHELL-5 paper describes a public in-car benchmark related to the earlier ICMC-ASR data. Do not assume the two names represent independent recordings that can be combined without overlap checks. The OpenSLR release documents speech splits, microphone channels and approximately 40 hours of separate noise. Noise hours are not transcribed speech hours.
Check English and European-language coverage by resource
The AMIE paper, published in 2020, describes passenger interactions collected during real rides. It is evidence of more recent English in-cabin research, not proof of a generally available commercial ASR training corpus. Its task and annotation scope should be checked against your own evaluation needs.
AVICAR and SPEECHDAT-CAR are older resources worth checking separately. Recording age alone does not prove that data is unusable, and a short list of public papers does not establish that no other English or European-language data exists. Verify language, microphone geometry, vehicle conditions, labels and access for the specific release under consideration.
For the wider application context, see the automotive voice data guide. For other public corpora, compare open speech datasets by the same requirements.
Match coverage to the deployment
Write down the languages and accents, seat positions, microphone layout, vehicle conditions and speaking tasks that matter. Then compare each candidate against that list. Identify whether it includes isolated commands, natural conversation, overlapping speakers, navigation requests or passenger intents; a language match alone is insufficient.
Keep training and evaluation partitions separate according to the agreed speaker and session rules. Inspect reference transcripts, speaker labels and channel maps before relying on a published benchmark score. An error rate from a different task or recording setup is not a forecast for your product.
Use augmentation as a measured experiment
Cabin noise recordings can support a controlled augmentation experiment with speech you are permitted to use. Record which noise sources, levels and transformations were applied. Compare the result against representative real-cabin recordings, using conditions and an SNR range chosen for the deployment rather than a universal threshold.
Adding noise alone does not establish coverage of microphone response, reverberation, speaker behavior in noise or playback echo. Evaluate the relevant conditions explicitly. A simulation may be useful without proving that every real-world gap is closed; equally, existing data should not be dismissed before testing it.
Our speech dataset catalog can be evaluated as a possible source of base speech. Confirm the selected asset's conditions and terms; catalog availability does not establish in-car capture.
Check the data terms, not just the paper
Access and permission are separate checks. Record the release identifier, download source and applicable terms, including attribution and sharing obligations. Review whether the proposed use is covered. A paper's publication license and a repository's software license are not substitutes for the dataset terms.
For AISHELL-5, use the license listed by its data distributor rather than describing the release as research-only. For ICMC-ASR, review the data user agreement associated with the supplied release. For AMIE or another research description, confirm whether data is available at all before planning around it.
Compare existing data, augmentation and collection
Use existing data when its coverage and terms fit the task. Consider an augmentation pilot where the remaining question is a testable acoustic gap. Consider custom collection when necessary language, recording or annotation conditions remain missing after those checks.
A small evaluation collection can help establish which gaps matter before a larger training-data commitment. Specify the vehicles, microphones, scenarios and acceptance checks, then confirm feasibility and timing in writing. Our in-car speech data collection guide explains the scoping questions. Collection is an option to assess, not a conclusion justified by an incomplete inventory.