Speech Technologies for a Multilingual World
summary
Crowdsourced speech dataset Common Voice curates 130+ languages (largest hours: Catalan), released quarterly under CC0 licence, built only after language communities request inclusion and with optional metadata tagging for demographics and recording conditions. Focus areas: read-speech vs natural and code-switched recordings, consent and copyright, severe bias by gender/age/accent, and on-device or privacy-preserving deployment. Practical takeaways: reuse free open corpora, collect conversational clips for real-world ASR, and expect nontrivial effort to cover regional variants and underrepresented speakers.
AI-written summary · about
chapters (13)
- 0:00Introduction and context
- 1:59Understanding speech tech
- 2:59The mechanics of ASR
- 5:31Benefits of voice technology
- 7:16Diverse applications
- 8:32The language gap
- 9:19Bias and representation
- 9:46Open-source data collection
- 19:08Accents and variants
- 20:19Data collection challenges
- 22:57Improving data quality
- 27:54Promoting accessibility
- 28:52Final thoughts and call to action
related talks
44mCommon Voice - Building Multilingual Voice Datasets
45mMore voice, less choice: The rise of voice interfaces and the decline of open source voice
46mArtificial intelligence to the rescue of accessibility
27mEnsuring Black voices matter: Why your voice assistant is racist, and what we can do about it
28mWhat’s Needed in Tech to bring Voice to the Future
21mMultimodal Open Source at Kyutai, From Online Demos to On-Device
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
58mAdding Speech to your Web App_UIArchitecture