MInDS-14 Multilingual Spoken Intent Detection

## Overview MInDS-14 is a multilingual spoken language understanding dataset for intent classification in the e-banking domain. It contains roughly 10K-100K spoken utterances across 14 language varieties (en-US, en-GB, en-AU, fr-FR, it-IT, es-ES, pt-PT, de-DE, nl-NL, ru-RU, pl-PL, cs-CZ, ko-KR, zh-CN), each labeled with one of 14 banking intents extracted from a commercial system. Data is distributed as parquet with audio and text modalities. ## Schema - path — string — file path to the audio clip - audio — audio — waveform (array + sampling_rate) - transcription — string — written form of the spoken utterance - english_transcription — string — English translation of the utterance - intent_class — int — label index (0-13) for the intent - lang_id — int — language variety identifier ## Sources - PolyAI/minds14 on Hugging Face — https://huggingface.co/datasets/PolyAI/minds14 — CC-BY-4.0 - Paper: Gerz et al., "Multilingual and Cross-Lingual Intent Detection from Spoken Data" (arXiv:2104.08524) ## Methodology Utterances were collected by crowd workers and expert annotators reading or paraphrasing scripted e-banking queries grounded in a commercial intent taxonomy. Each language variety was recorded independently to reflect natural speech in that locale. Annotations combine expert-generated, crowdsourced, and machine-generated labels per the source card. Audio is provided as raw waveforms suitable for resampling to common ASR rates. ## Known gaps & limitations Coverage is restricted to 14 banking intents — not a general-purpose SLU benchmark. Per-language splits are small (hundreds to low thousands of clips), which limits statistical power for fine-grained evaluation. Speaker demographics, microphone conditions, and dialectal coverage within each language variety are not exhaustively documented. The dataset is widely used in published benchmarks, so leakage risk exists when training models intended for the same evaluations. ## Intended use & out-of-scope - Intended: benchmarking multilingual spoken intent classification, evaluating ASR+NLU pipelines, few-shot/cross-lingual transfer research, audio-text alignment evals for speech LLMs. - Out-of-scope: training production banking voice assistants without domain adaptation; general open-domain SLU; speaker identification or biometrics. _Federated dataset: 16 parquet shards, 1.06 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: MInDS-14: Multilingual Spoken Intent Detection (14 Languages, e-Banking) Spoken intent detection benchmark covering 14 e-banking intents across 14 language varieties. Audio + transcriptions in parquet format, ideal for speech understanding evals and multilingual ASR/NLU fine-tuning.