Scaling Semi-Supervised Production-Grade ASR on 200 Languages
summary
Describes a concrete pipeline that scales semi supervised, self supervised ASR to 200 languages using Wav2Vec2/XLSR style pretraining and downstream fine tuning, reports pretraining on ~400k hours of audio and experiments with 300M–1B parameter models, lays out data acquisition, base‑model selection, heavy human‑in‑the‑loop tuning, engineering optimizations, and deployment choices, and concludes the approach is feasible for production but costly in compute and annotation and requires careful hyperparameter and inference engineering.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 4:25ASR pipeline components
- 9:29Data's role in ASR
- 14:44Base model estimation
- 21:51Self-supervised pretraining
- 23:17Human backprop tuning
- 26:50Trend validate tune
- 27:34Transformers vs Conformers
- 28:58Loss functions for ASR
- 30:03Inference optimization
- 33:14ASR deployment best practices
- 37:34Scaling ASR advice
- 38:16Q&A
related talks
31mBuilding Large-scale Localized Language Models
20mMeta's AI Translations Inference
53mSelf-Supervised Learning in Transformer Models for Birdsong
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
25mTraining Large Language Models in the Open
16mThe End of Awkward AI Transcriptions
30mProductionizing an unsupervised machine learning model to...
01h 38mTraining Large Language Models: Practices and Research Questions