Source-linked AI summary

FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech

Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, Ankur Bapna

arXiv:2205.12446v1cs.CLcs.LGcs.SDeess.AS

TL;DR

FLEURS addresses the limited availability of broad multilingual speech evaluation by introducing an n-way parallel dataset in 102 languages. It constructs the resource from FLoRes-101 and evaluates multilingual pre-trained representations across recognition, identification, translation, and retrieval tasks. The paper concludes that FLEURS is suited to data-efficient multilingual evaluation and reports baseline performance across its tasks.

  • Problem

    Existing multilingual speech datasets cover fewer languages or narrower task settings, limiting evaluation across more than 100 languages.

  • Method

    FLEURS builds n-way parallel speech and text from FLoRes-101 and fine-tunes multilingual pre-trained models for several speech tasks.

  • Results

    The paper provides baseline ASR, Speech LangID, and retrieval systems and reports 73.3% macro-average accuracy for mSLAM on FLEURS LangID.

  • Takeaways & Limitations

    FLEURS is presented as a benchmark for evaluating data-efficient multilingual speech and text representations across 102 languages.

  • Takeaways & Limitations

    FLEURS-LangID uses different speakers across train and dev/test, making avoidance of speaker-ID overfitting essential for good performance.

Abstract

from arXiv · show

We introduce FLEURS, the Few-shot Learning Evaluation of Universal Representations of Speech benchmark. FLEURS is an n-way parallel speech dataset in 102 languages built on top of the machine translation FLoRes-101 benchmark, with approximately 12 hours of speech supervision per language. FLEURS can be used for a variety of speech tasks, including Automatic Speech Recognition (ASR), Speech Language Identification (Speech LangID), Translation and Retrieval. In this paper, we provide baselines for the tasks based on multilingual pre-trained models like mSLAM. The goal of FLEURS is to enable speech technology in more languages and catalyze research in low-resource speech understanding.

1. Introduction

FLEURS addresses the need for massively multilingual speech evaluation by providing n-way parallel speech and text in 102 languages. It supports diverse speech tasks and includes baselines based on multilingual pre-trained models.

  • Unlike other cited parallel speech datasets, FLEURS spans over 100 languages and supports diverse language research.
  • FLEURS is an n-way parallel speech and text dataset covering 102 languages.
  • The dataset provides natural human speech, high-quality transcripts, and strong quality control.
  • FLEURS uses bottom-up collection of spoken utterances for aligned segments rather than primarily document-level automatic alignment.
  • FLEURS supports ASR, speech translation, Speech LangID, and multilingual speech-text retrieval.
  • The paper provides Speech LangID, ASR, and speech-text retrieval baselines by fine-tuning mSLAM and multilingual w2v-BERT.

2. Dataset

FLEURS builds a multilingual speech dataset from parallel FLoRes-101 sentences, with recordings from native speakers and transcripts inherited from human translations. The dataset preserves parallel indexing, linguistic diversity, and reproducibility-oriented preprocessing.

  • Speech Data Collection: FLEURS is constructed from 2009 FLoRes-101 dev and devtest sentences across 102 languages.The splits contain 1509 training, 150 development, and 350 test sentences.
  • Speech Data Collection: Each sentence initially receives three recordings from three native speakers, with an attempted sex-ratio balance of at least 30/70%.
  • Speech Data Collection: About 21.5% of sentences are missing in the first dataset version because none of their recordings were validated.The authors plan to fill these gaps in future versions.
  • Transcripts and Representation: FLEURS reuses human-translated source transcripts and retains English translations for multimodal speech-translation evaluation.
  • Transcripts and Representation: The dataset adopts a universal character vocabulary as both its modeling unit and character-level evaluation unit.
  • Taxonomy and Statistics: FLEURS covers 17 language families and 27 writing systems, with sentence indices enabling recovery of n-way parallelism.

3. Task Baselines

The baselines use pre-training followed by fine-tuning because training a giant model directly on FLEURS would overfit. The setup initializes multilingual fine-tuning from a large wav2vec-BERT model trained on unlabeled speech.

  • Experimental Setup: The baseline methodology fine-tunes multilingual pre-trained models rather than training a giant model from scratch on FLEURS.The paper states that direct training would easily overfit.
  • Experimental Setup: The main initialization is a 600M-parameter wav2vec-BERT model pre-trained on 429k hours of unlabeled speech in 51 languages.
  • Experimental Setup: The baselines cover speech recognition, Speech LangID, and speech-text retrieval tasks.
  • Language Splits: The experiments distinguish 54 seen languages, whose speech data appeared in pre-training, from 48 unseen languages without such speech data.

4. Downstream Tasks

FLEURS is evaluated across multilingual ASR, Speech LangID, and cross-modal speech-text retrieval using pretrained-model baselines. Performance varies substantially by language group and pretraining coverage, with consistent challenges for underrepresented or unseen languages.

  • Speech recognition: 10.7 and 10.6 average CER were reported for Western European languages, compared with 15.6 and 16.4 for Sub-Saharan African languages and 17.4 and 19.2 for South Asian languages.The paper partly attributes these gaps to larger amounts of unlabeled Western European pretraining data.
  • Speech recognition: Speech-only pretraining produced 0.5% lower CER than speech-text pretraining across the multilingual ASR baselines.The regression with speech-text initialization was observed especially in South Asian, Sub-Saharan African, and CJK groups.
  • Speech recognition: Languages unseen during pretraining often had worse-than-average test CER because fine-tuning used very limited supervised data.Some unseen languages nevertheless achieved below-average CER, including languages using Latin, Cyrillic, Malayalam, Kannada, Gujarati, Devanagari, Burmese, and Arabic scripts.
  • Speech Language Identification: 73.3% macro-average accuracy was obtained by mSLAM on FLEURS Speech LangID, versus 71.4% for multilingual w2v-bert-51.Group accuracy decreased in the order CJK, Western European, Eastern European, Central-Asia/Middle-East/North-Africa, South-East Asia, Sub-Saharan Africa, and South Asia.
  • Cross-modal Speech-Text Retrieval: 76.9% P@1 was achieved for speech-to-text retrieval and 74.4% P@1 for text-to-speech retrieval.Seen languages generally outperformed unseen languages, with especially steep degradation for CJK languages.
  • Cross-modal Speech-Text Retrieval: Retrieval performance exposed modality- and script-specific weaknesses: Odia lacked tokenizer representation, while Urdu performed considerably worse for text-to-speech retrieval than speech-to-text retrieval.The paper associates Urdu’s difficulty with phonetic similarity to other South Asian languages and missing speech-modality pretraining.

5. Conclusion

FLEURS is a 102-language n-way parallel speech dataset for evaluating speech recognition, translation, classification, and retrieval methods, especially data-efficient multilingual pre-trained representations.

  • FLEURS is an n-way parallel speech dataset covering 102 languages.
  • FLEURS supports evaluation of speech recognition, translation, classification, and retrieval methods.
  • Baseline ASR, language identification, and retrieval systems show FLEURS is suited to evaluating data-efficient multilingual pre-trained speech and text representations.
  • The dataset is intended to catalyze few-shot understanding research across many languages.
Loading 2205.12446v1…