Source-linked AI summary
Common Voice: A Massively-Multilingual Speech Corpus
Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M. Tyers, Gregor Weber
TL;DR
Common Voice addresses the scarcity and cost of speech-training data for many languages by building an open, community-scalable corpus through crowdsourced recording and validation. The paper demonstrates multilingual ASR transfer learning with the corpus and presents it as a large public-domain resource designed for ASR.
Problem
Speech-technology training data is prohibitively expensive or unavailable for most languages, limiting accessible multilingual research resources.
Method
Common Voice uses community contributions to collect and validate speech, while multilingual ASR experiments transfer layers from a pretrained English model and fine-tune them for target languages.
Results
Common Voice is presented as the largest public-domain corpus designed for Automatic Speech Recognition, with 38 language communities collecting data.
Takeaways & Limitations
The corpus provides an open, scalable resource for collecting speech in both minority and majority languages and for multilingual ASR research.
Takeaways & Limitations
The reported experiments used an internal February 2019 corpus version whose train, development, and test splits do not exactly match official releases.
Abstract
from arXiv · showhide
The Common Voice corpus is a massively-multilingual collection of transcribed speech intended for speech technology research and development. Common Voice is designed for Automatic Speech Recognition purposes but can be useful in other domains (e.g. language identification). To achieve scale and sustainability, the Common Voice project employs crowdsourcing for both data collection and data validation. The most recent release includes 29 languages, and as of November 2019 there are a total of 38 languages collecting data. Over 50,000 individuals have participated so far, resulting in 2,500 hours of collected audio. To our knowledge this is the largest audio corpus in the public domain for speech recognition, both in terms of number of hours and number of languages. As an example use case for Common Voice, we present speech recognition experiments using Mozilla's DeepSpeech Speech-to-Text toolkit. By applying transfer learning from a source English model, we find an average Character Error Rate improvement of 5.99 +/- 5.48 for twelve target languages (German, French, Italian, Turkish, Catalan, Slovenian, Welsh, Irish, Breton, Tatar, Chuvash, and Kabyle). For most of these languages, these are the first ever published results on end-to-end Automatic Speech Recognition.
1. Introduction
Common Voice addresses the lack of affordable or available speech-technology training data for many languages through community-driven, open, and scalable corpus construction.
- Common Voice targets languages for which speech-technology training data is prohibitively expensive or unavailable.
- The project combines community building, open-source tooling, and permissive licensing to support open and decentralized speech technology.
- Communities can scale the corpus to new languages by translating the interface, submitting text sentences, and recording and validating voices.
2. Prior work
Prior multilingual speech corpora provide useful alternatives but differ from Common Voice in licensing, sustainability, validation, or community participation.
- Babel contains high-quality data from 22 minority languages but is not released under an open license.
- VoxForge is community-driven, multilingual across 17 languages, and openly licensed, but lacks a sustainable collection pipeline and data validation.
- M-AILABS provides nine language varieties under a modified BSD 3-Clause License but lacks a community-driven aspect.
- Common Voice is presented as a sustainable, open alternative supporting collection of both minority and majority languages.
3. Corpus Creation
Common Voice collects speech through contributor recordings and validates clips through independent contributor votes, while separating speakers and repeated sentences across dataset splits.
- Contributors record displayed sentences through the Common Voice website or iPhone app, and other contributors later verify the recordings.
- Clips become valid after two up-votes or invalid after two down-votes, with at most three contributors listening to any clip.
- Early voting bugs caused some official-release clips to receive more than three votes, which are handled using a simple majority rule.
- Any speaker appears in only one split, and repeated text sentences are removed from the train, test, and development sets.
- Dataset split sizes are determined using statistical power analyses targeting a 99% confidence level and 1% margin of error relative to training-set clips.
- Released audio uses mono-channel, 16-bit MPEG-3 files sampled at 48kHz because the web platform requires broadly supported recording and playback.
4. Corpus Contents
Common Voice distributes language-specific datasets with audio, transcripts, vote information, and optional demographics, and adds languages through community translation and locally gathered prompts.
- 4.1. Released Languages: Each language download contains six TSV files and an audio directory, with columns for speaker ID, file path, sentence, votes, age, gender, and accent.
- 4.1. Released Languages: Speaker identifiers are anonymized, while demographic fields are optionally self-reported.
- 4.2. Adding a new Language: Adding a language requires translating the web interface and gathering text prompts to be read aloud.
- 4.2. Adding a new Language: The interface contains 610 text strings translated through the community-managed Pontoon platform.
- 4.2. Adding a new Language: Communities extract initial prompts from Wikipedia with language-specific rules and can submit additional sentences through the Sentence Collector.
- 4.2. Adding a new Language: Every Sentence Collector submission requires approval from two of three reviewers before weekly export into Common Voice.
5. Automatic Speech Recognition Experiments
The experiments evaluate Common Voice for multilingual end-to-end speech recognition using transfer learning from a pretrained English DeepSpeech model. They use an earlier multilingual dataset version with speaker-independent splits and a six-layer CTC architecture.
- Experimental approach: Transfer learning copies selected layers from a pretrained English source model, initializes target-language layers, stitches the networks, and trains all layers jointly.The approach is intended to avoid requiring linguistic resources or domain expertise.
- Data: Speaker-disjoint splits keep each speaker’s recordings in only one partition, enabling evaluation of speaker generalization.Splits were approximately 80% training, 10% development, and 10% test.
- Model architecture: The DeepSpeech v0.3.0 model is a six-layer end-to-end ASR network trained with CTC loss, containing three fully connected layers, one unidirectional LSTM, and two further fully connected layers.The architecture uses 2,048-dimensional hidden layers and target-language character outputs.
- Data: The dataset uses an earlier multilingual Common Voice version with language-specific numbers of audio clips and unique speakers.The supplied table caption identifies these two quantities as the reported dataset statistics.
- Training: Training uses batch sizes of 24 for training and 48 for development, 20% dropout, a 0.0001 learning rate, ADAM, Xavier initialization, and development-loss early stopping.Early stopping is triggered by specified development-loss behavior over five-epoch windows.
6. Results
The results compare fine-tuned transfer-learning models with target-language models trained from scratch across languages and transferred-layer counts. Four copied layers produce the best final model trend across the investigated languages.
- Per-language results: Table 3 reports Character Error Rate for each language and includes a baseline trained from scratch on target-language data.Bold values identify the best model per language, while darker shading indicates more accurate relative performance.
- Transfer-learning results: Four layers copied from pretrained English DeepSpeech produce the best final model trend across the target languages.The trend is reported from the per-language Character Error Rate results and becomes clearer when averaged across languages.
7. Concluding remarks
The paper presents Common Voice as a crowd-sourced multilingual speech corpus designed to scale through community effort and released under a permissive public license. It also reports multilingual ASR experiments using the corpus and highlights Character Error Rate as a language-agnostic metric.
- Corpus contribution: Common Voice is a crowd-sourced multilingual speech corpus intended to scale to any language through community effort.The project uses volunteer contributions to support its continuing expansion.
- Corpus contribution: All speech data are released under a Creative Commons CC0 license.This makes the corpus openly available for research and development.
- Evaluation: Character Error Rate is used instead of Word Error Rate because it is more language-agnostic in multilingual evaluation.The paper specifically notes that CER is more appropriate for languages without clearly delimited words.
- Experiments: The paper presents multilingual end-to-end ASR experiments using Common Voice and describes the corpus as the largest public-domain corpus designed for ASR.The conclusion notes 38 language communities were collecting data at the time of writing.