MInDS-14: A Multilingual Spoken Intent Dataset

MInDS-14: A Multilingual Spoken Intent Dataset

TL;DR

MInDS-14 is a multilingual spoken intent dataset built for training and evaluating voice intent classifiers. It includes 8,168 labeled audio samples across 14 languages and 14 e-banking intent classes.

Now listed on AIOZ AI, MInDS-14 helps developers build multilingual voice assistant workflows without recording, transcribing, translating, and labeling audio data from scratch.

What MInDS-14 Is

MInDS-14 is a speech dataset designed for multilingual spoken intent detection.

Each sample contains a spoken customer request, the original-language transcription, an English translation, and an intent label. The dataset focuses on e-banking scenarios, making it useful for voice assistants, IVR systems, and speech classification models that need to route customer requests by intent.

For builders working with models such as wav2vec2, Whisper, or HuBERT, MInDS-14 provides a structured starting point for multilingual speech classification.

How the Dataset Is Organized

MInDS-14 includes audio samples across 14 languages and 14 banking-related intent classes.

Each record includes:

  • Audio: the spoken utterance
  • Transcription: text in the original language
  • English transcription: translated version of the utterance
  • Intent class: one of 14 e-banking intents
  • Language ID: the language label for the sample

Intent classes include balance checks, address changes, card issues, cash deposits, direct debit, high-value payments, latest transactions, pay bills, and related customer service requests.

Key Dataset Details

  • Dataset: MInDS-14
  • Dataset type: multilingual spoken intent dataset
  • Total examples: 8,168
  • Languages: 14
  • Intent classes: 14
  • Domain: e-banking
  • Splits: train only
  • Audio sampling rate: 8000 Hz
  • Audio duration range: 1.37 to 67.9 seconds
  • Download size: ~628 MB
  • License: CC BY 4.0

Where It Fits Best

  • Multilingual voice banking assistants
  • Spoken intent classifiers for IVR systems
  • Cross-lingual transfer learning experiments
  • Speech model benchmarking
  • Audio preprocessing and resampling workflows
  • Fine-tuning speech models for intent classification

Access MInDS-14 on AIOZ AI

MInDS-14 gives builders labeled spoken audio across 14 languages and 14 e-banking intents.

Fourteen languages, one dataset for multilingual voice intent detection.

Access MInDS-14 on AIOZ AI and start building spoken intent workflows without starting from scratch.

FAQ

Q1: What is MInDS-14 used for?

It is used to train and evaluate multilingual spoken intent classification models, especially for e-banking voice assistant workflows.

Q2: How many examples does MInDS-14 include?

It includes 8,168 labeled audio samples.

Q3: What should builders watch for before fine-tuning?

The audio is recorded at 8000 Hz, while many pretrained speech models expect 16000 Hz, so upsampling is usually needed before fine-tuning.