AI People Wiki

Builders & open source

Boris Ginsburg

Applied research scientist, NVIDIA

  • Applied research scientist, NVIDIA1

Boris Ginsburg is an applied research scientist at NVIDIA whose work spans fast, scalable training algorithms for neural networks and speech recognition. Ginsburg is the last-listed author of the papers behind NVIDIA's Parakeet models, including the Token-and-Duration Transducer, the FastConformer encoder, the Parakeet-TDT-0.6B-v3 technical report and the unified streaming transducer used by parakeet-unified-en-0.6b.12345

At a glance

Current role
Applied research scientist
Company
NVIDIA
Sector
Company
Work listed
3
Sources cited
12
Profile reviewed

Training at scale

NVIDIA's author page describes Ginsburg as an applied research scientist whose primary research is fast and scalable algorithms for training neural networks. Ginsburg co-authored "Large Batch Training of Convolutional Networks" (2017), which introduced layer-wise adaptive rate scaling, and "Mixed Precision Training" (2017).167

Ginsburg's NVIDIA technical blog posts cover mixed precision training for NLP and speech with OpenSeq2Seq (2018), pretraining BERT with layer-wise adaptive learning rates (2019), and building smaller speech recognition models with NVIDIA NeMo (2019).1

Speech recognition

Ginsburg co-authored the Jasper (2019) and QuartzNet (2019) convolutional acoustic models and the 2019 paper introducing NeMo, a toolkit for building AI applications from reusable neural modules with collections for speech recognition and language processing.8910

Ginsburg is the last-listed author of "Efficient Sequence Transduction by Jointly Predicting Tokens and Durations" (2023) and "Fast Conformer with Linearly Scalable Attention" (2023), both cited by the Parakeet-TDT-0.6B-v3 model card, of the 2025 Canary-1B-v2 and Parakeet-TDT-0.6B-v3 technical report, and of the 2026 unified offline-streaming transducer paper cited by the parakeet-unified-en-0.6b model card.23451112

Work and contributions

  • NVIDIA parakeet-tdt-0.6b-v3 Open weights

    Last-listed author of the 2025 technical report introducing Canary-1B-v2 and Parakeet-TDT-0.6B-v3, which the model card points to for architecture and training details.411

  • NVIDIA parakeet-unified-en-0.6b Open weights

    Last-listed author of the 2026 paper on a unified offline and streaming transducer framework, the first reference in the parakeet-unified-en-0.6b model card.512

  • NVIDIA Parakeet Open weights

    Last-listed author of the Token-and-Duration Transducer (TDT) and FastConformer papers, both cited as references in the Parakeet-TDT-0.6B-v3 model card.2311

Timeline

  1. Co-authored "Large Batch Training of Convolutional Networks" with Yang You and Igor Gitman.6
  2. Co-authored the "Mixed Precision Training" paper.7
  3. Co-authored Jasper, an end-to-end convolutional neural acoustic model.8
  4. Co-authored the paper introducing the NeMo toolkit for building AI applications from neural modules.10
  5. Co-authored QuartzNet, an ASR model built from 1D time-channel separable convolutions.9
  6. Last-listed author of the Token-and-Duration Transducer (TDT) paper.2
  7. Last-listed author of the FastConformer paper on linearly scalable attention for speech recognition.3
  8. Last-listed author of the Canary-1B-v2 and Parakeet-TDT-0.6B-v3 technical report.4
  9. Last-listed author of the unified offline-streaming transducer paper behind parakeet-unified-en-0.6b.5

Where they work

Suggest a correction or removal

Profiles cover public professional work only, and each statement cites a source. If something is out of date, missing or wrong, tell us and link a source. People can also ask for their profile to be corrected or removed.

References

12 sources · profile reviewed Oct 22, 2026 · data updated Oct 22, 2026

  1. Author: Boris Ginsburg (NVIDIA Technical Blog)NVIDIA · Official page · accessed Oct 22, 2026
  2. Efficient Sequence Transduction by Jointly Predicting Tokens and Durations (arXiv 2304.06795)arXiv · Paper · published Apr 13, 2023 · accessed Oct 22, 2026
  3. Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition (arXiv 2305.05084)arXiv · Paper · published May 8, 2023 · accessed Oct 22, 2026
  4. Canary-1B-v2 & Parakeet-TDT-0.6B-v3: Efficient and High-Performance Models for Multilingual ASR and AST (arXiv 2509.14128)arXiv · Paper · published Sep 17, 2025 · accessed Oct 22, 2026
  5. Reducing the Offline-Streaming Gap for Unified ASR Transducer with Consistency Regularization (arXiv 2604.19079)arXiv · Paper · published Apr 21, 2026 · accessed Oct 22, 2026
  6. Large Batch Training of Convolutional Networks (arXiv 1708.03888)arXiv · Paper · published Aug 13, 2017 · accessed Oct 22, 2026
  7. Mixed Precision Training (arXiv 1710.03740)arXiv · Paper · published Oct 10, 2017 · accessed Oct 22, 2026
  8. Jasper: An End-to-End Convolutional Neural Acoustic Model (arXiv 1904.03288)arXiv · Paper · published Apr 5, 2019 · accessed Oct 22, 2026
  9. QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions (arXiv 1910.10261)arXiv · Paper · published Oct 22, 2019 · accessed Oct 22, 2026
  10. NeMo: a toolkit for building AI applications using Neural Modules (arXiv 1909.09577)arXiv · Paper · published Sep 14, 2019 · accessed Oct 22, 2026
  11. nvidia/parakeet-tdt-0.6b-v3 model cardNVIDIA (Hugging Face) · Model card · accessed Oct 22, 2026
  12. nvidia/parakeet-unified-en-0.6b model cardNVIDIA (Hugging Face) · Model card · accessed Oct 22, 2026