Uploaded October 2024 | Updated September 2026, 1 week ago
Will Williams is CTO of Speechmatics in Cambridge. In this sponsored episode - he shares deep technical insights into modern speech recognition technology and system architecture. The episode covers several key technical areas:
Will Williams is CTO of Speechmatics, the Cambridge-based speech recognition company. Tim has used their API for years to caption MLST episodes, and this conversation happened in their offices -- which explains the live demo at the start where an AI moderates a political debate clip in real time.
The technical meat covers how Speechmatics builds production ASR systems. Their approach is hybrid: self-supervised pre-training on unlabeled audio gets them comparable accuracy to fully supervised systems like Whisper, but with roughly 100x less labeled data. Williams explains why this matters for scaling to low-resource languages where you simply don't have thousands of hours of human-transcribed speech.
The architecture discussion is detailed. Their system runs multiple operating points with different latency-accuracy tradeoffs. They pad latency up to 1.8 seconds to keep the user experience consistent rather than optimizing for raw speed. Decoding uses lattices with language model integration, which lets them rescore hypotheses and handle things like proper nouns and domain-specific vocabulary without retraining the acoustic model.
Diarization -- figuring out who said what -- comes up repeatedly. Williams calls it harder than ASR itself, partly because speaker embeddings get corrupted by acoustic environments and partly because cross-talk creates genuinely ambiguous boundaries. They're pushing hard on implicit source separation but the problem remains open.
The conversation also covers their testing infrastructure (mirrored production traffic catches edge cases that unit tests miss), why they resist customer-specific fine-tuning (it fragments the model and makes global improvements harder), and Williams' critique of PyTorch memory management in production settings. He argues for more direct memory allocation rather than letting the framework handle it, which is a practical concern when you're serving models at scale.
Featuring: Will Williams (CTO, Speechmatics) and Tim Scarfe.
---
TIMESTAMPS:
00:00:00 ASR and diarization fundamentals
00:05:25 Real-time conversational AI architecture
00:09:21 Neural network streaming and multi-modal integration
00:12:49 Enterprise voice AI and real-time translation
00:20:00 Production deployment and testing infrastructure
00:29:38 Model architecture and latency-accuracy tradeoffs
00:35:40 Lattice-based decoding and language model integration
00:44:00 ASR performance metrics and real-world evaluation
00:51:30 Ethics and privacy in speech technology
01:00:50 Self-supervised learning and low-resource languages
01:11:00 Feature engineering to automated ML
01:21:00 Infrastructure scaling and PyTorch critique
01:35:00 Future of conversational AI and Ursa 2
---
REFERENCES:
paper:
[00:00:05] Speechmatics PDF shownotes
dropbox.com/scl/fi/d94b1jcgph9o8au8shdym/Speechmatics.pdf?rlkey=bi55wvktzomzx0y5sic6jz99y&st=6qwofv8t&dl=0
[00:10:09] GFlowNets
arxiv.org/abs/2106.04399
[01:35:00] Ursa 2 model
speechmatics.com/company/articles-and-news/ursa-2-elevating-speech-recognition-across-52-languages
company:
[00:01:15] Speechmatics
speechmatics.com
person:
[00:01:32] Will Williams
https://x.com/wjwwilliams
---
LINKS:
Full Transcript: app.rescript.info/share/c6887b6d7b214f93daad1c18d70e2eb6
Download PDF transcript: app.rescript.info/api/public/sessions/abeef42b31287680/pdf
Will Williams, CTO, Speechmatics
https://x.com/wjwwilliams
Will Williams is CTO of Speechmatics in Cambridge. In this sponsored episode - he shares deep technical insights into modern speech recognition technology and system architecture. The episode covers several key technical areas:
Will Williams is CTO of Speechmatics, the Cambridge-based speech recognition company. Tim has used their API for years to caption MLST episodes, and this conversation happened in their offices -- which explains the live demo at the start where an AI moderates a political debate clip in real time.
The technical meat covers how Speechmatics builds production ASR systems. Their approach is hybrid: self-supervised pre-training on unlabeled audio gets them comparable accuracy to fully supervised systems like Whisper, but with roughly 100x less labeled data. Williams explains why this matters for scaling to low-resource languages where you simply don't have thousands of hours of human-transcribed speech.
The architecture discussion is detailed. Their system runs multiple operating points with different latency-accuracy tradeoffs. They pad latency up to 1.8 seconds to keep the user experience consistent rather than optimizing for raw speed. Decoding uses lattices with language model integration, which lets them rescore hypotheses and handle things like proper nouns and domain-specific vocabulary without retraining the acoustic model.
Diarization -- figuring out who said what -- comes up repeatedly. Williams calls it harder than ASR itself, partly because speaker embeddings get corrupted by acoustic environments and partly because cross-talk creates genuinely ambiguous boundaries. They're pushing hard on implicit source separation but the problem remains open.
The conversation also covers their testing infrastructure (mirrored production traffic catches edge cases that unit tests miss), why they resist customer-specific fine-tuning (it fragments the model and makes global improvements harder), and Williams' critique of PyTorch memory management in production settings. He argues for more direct memory allocation rather than letting the framework handle it, which is a practical concern when you're serving models at scale.
Featuring: Will Williams (CTO, Speechmatics) and Tim Scarfe.
---
TIMESTAMPS:
00:00:00 ASR and diarization fundamentals
00:05:25 Real-time conversational AI architecture
00:09:21 Neural network streaming and multi-modal integration
00:12:49 Enterprise voice AI and real-time translation
00:20:00 Production deployment and testing infrastructure
00:29:38 Model architecture and latency-accuracy tradeoffs
00:35:40 Lattice-based decoding and language model integration
00:44:00 ASR performance metrics and real-world evaluation
00:51:30 Ethics and privacy in speech technology
01:00:50 Self-supervised learning and low-resource languages
01:11:00 Feature engineering to automated ML
01:21:00 Infrastructure scaling and PyTorch critique
01:35:00 Future of conversational AI and Ursa 2
---
REFERENCES:
paper:
[00:00:05] Speechmatics PDF shownotes
dropbox.com/scl/fi/d94b1jcgph9o8au8shdym/Speechmatics.pdf?rlkey=bi55wvktzomzx0y5sic6jz99y&st=6qwofv8t&dl=0
[00:10:09] GFlowNets
arxiv.org/abs/2106.04399
[01:35:00] Ursa 2 model
speechmatics.com/company/articles-and-news/ursa-2-elevating-speech-recognition-across-52-languages
company:
[00:01:15] Speechmatics
speechmatics.com
person:
[00:01:32] Will Williams
https://x.com/wjwwilliams
---
LINKS:
Full Transcript: app.rescript.info/share/c6887b6d7b214f93daad1c18d70e2eb6
Download PDF transcript: app.rescript.info/api/public/sessions/abeef42b31287680/pdf
Will Williams, CTO, Speechmatics
https://x.com/wjwwilliams
![Neural Networks Are Elastic Origami! [Prof. Randall Balestriero]](https://i.ytimg.com/vi/l3O2J3LMxqI/mqdefault.jpg "Neural Networks Are Elastic Origami! [Prof. Randall Balestriero]
Professor Randall Balestriero joins us to discuss neural network geometry, spline theory, and emerging phenomena in deep learning, based on research presented at ICML. Topics include the delayed emergence of adversarial robustness in neural networks (grokking), geometric interpretations of neural networks via spline theory, and challenges in reconstruction learning. We also cover geometric analysis of Large Language Models (LLMs) for toxicity detection and the relationship between intrinsic dimensionality and model control in RLHF.
SPONSOR MESSAGES:
***
CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments.
https://centml.ai/pricing/
Tufa AI Labs is a brand new research lab in Zurich started by Benjamin Crouzier focussed on o-series style reasoning and AGI. Are you interested in working on reasoning, or getting involved in their events?
Goto https://tufalabs.ai/
***
Show notes and transcript: https://www.dropbox.com/scl/fi/3lufge4upq5gy0ug75j4a/RANDALLSHOW.pdf?rlkey=nbemgpa0jhawt1e86rx7372e4&dl=0
TOC:
[00:00:00] Introduction
1. Neural Network Geometry and Spline Theory
[00:01:41] 1.1 Neural Network Geometry and Spline Theory
[00:07:41] 1.2 Deep Networks Always Grok
[00:11:39] 1.3 Grokking and Adversarial Robustness
[00:16:09] 1.4 Double Descent and Catastrophic Forgetting
2. Reconstruction Learning
[00:18:49] 2.1 Reconstruction Learning
[00:24:15] 2.2 Frequency Bias in Neural Networks
3. Geometric Analysis of Neural Networks
[00:29:02] 3.1 Geometric Analysis of Neural Networks
[00:34:41] 3.2 Adversarial Examples and Region Concentration
4. LLM Safety and Geometric Analysis
[00:40:05] 4.1 LLM Safety and Geometric Analysis
[00:46:11] 4.2 Toxicity Detection in LLMs
[00:52:24] 4.3 Intrinsic Dimensionality and Model Control
[00:58:07] 4.4 RLHF and High-Dimensional Spaces
5. Conclusion
[01:02:13] 5.1 Neural Tangent Kernel
[01:08:07] 5.2 Conclusion
REFS:
[00:01:35] Balestriero/Humayun – Deep network geometry & input space partitioning
https://arxiv.org/html/2408.04809v1
[00:03:55] Balestriero & Paris – Linking deep networks to adaptive spline operators
https://proceedings.mlr.press/v80/balestriero18b/balestriero18b.pdf
[00:13:55] Song et al. – Gradient-based white-box adversarial attacks
https://arxiv.org/abs/2012.14965
[00:16:05] Humayun, Balestriero & Baraniuk – Grokking phenomenon & emergent robustness
https://arxiv.org/abs/2402.15555
[00:18:25] Humayun – Training dynamics & double descent via linear region evolution
https://arxiv.org/abs/2310.12977
[00:20:15] Balestriero – Power diagram partitions in DNN decision boundaries
https://arxiv.org/abs/1905.08443
[00:23:00] Frankle & Carbin – Lottery Ticket Hypothesis for network pruning
https://arxiv.org/abs/1803.03635
[00:24:00] Belkin et al. – Double descent phenomenon in modern ML
https://arxiv.org/abs/1812.11118
[00:25:55] Balestriero et al. – Batch normalization’s regularization effects
https://arxiv.org/pdf/2209.14778
[00:29:35] EU – EU AI Act 2024 with compute restrictions
https://www.lw.com/admin/upload/SiteAttachments/EU-AI-Act-Navigating-a-Brave-New-World.pdf
[00:39:30] Humayun, Balestriero & Baraniuk – SplineCam: Visualizing deep network geometry
https://openaccess.thecvf.com/content/CVPR2023/papers/Humayun_SplineCam_Exact_Visualization_and_Characterization_of_Deep_Network_Geometry_and_CVPR_2023_paper.pdf
[00:40:40] Carlini – Trade-offs between adversarial robustness and accuracy
https://arxiv.org/abs/1902.06705
[00:44:55] Balestriero & LeCun – Limitations of reconstruction-based learning methods
https://raw.githubusercontent.com/mlresearch/v235/main/assets/balestriero24b/balestriero24b.pdf
[00:47:20] Balestriero & LeCun – Spectral analysis of neural network learning
https://proceedings.neurips.cc/paper_files/paper/2022/file/aa56c74513a5e35768a11f4e82dd7ffb-Paper-Conference.pdf
[00:49:45] He et al. – MAE: Masked Autoencoders for self-supervised learning
https://arxiv.org/abs/2111.06377
[00:54:50] Balestriero et al. – Geometric analysis of LLM layers for toxicity detection
https://arxiv.org/abs/2309.12312
[00:59:35] Balestriero et al. – Superior toxicity detection via geometric features
https://arxiv.org/html/2312.01648v2
[01:04:45] UofT ML – Self-attention control & context length effects
https://arxiv.org/abs/2310.04444
[01:11:55] Roberts – Foundations of deep learning theory
https://arxiv.org/abs/2106.10165
[01:15:40] Balestriero & Cha – Kolmogorov GAM Networks via spline partition theory
https://arxiv.org/pdf/2501.00704
[01:16:40] Various – Graph Kolmogorov-Arnold Networks (GKAN) extension
https://www.nature.com/articles/s41598-024-85083-8")
![The ARC Prize 2024 Winning Algorithm [Daniel Franzen and Jan Disselhoff]](https://i.ytimg.com/vi/mTX_sAq--zY/mqdefault.jpg "The ARC Prize 2024 Winning Algorithm [Daniel Franzen and Jan Disselhoff]
SPONSOR MESSAGES:
***
CentML offers competitive pricing for GenAI model deployment, with flexible options to suit a wide range of models, from small to large-scale deployments. Check out their super fast DeepSeek R1 hosting!
https://centml.ai/pricing/
Daniel Franzen and Jan Disselhoff the ARChitects won the ARC Prize 2024 with co-researcher David Hartmann, achieving a remarkable 53.5% accuracy on the ARC challenge using a 12-billion parameter language model. Filmed at Tufa Labs in Zurich, they walk through their solution architecture in detail: how they tokenized grid-based visual puzzles as text and fed them directly into an LLM, why test-time training on evaluation examples gave a major score boost, and how depth-first search over token probabilities outperformed greedy and beam search for solution generation.
The conversation gets into the surprising computational capabilities of language models on spatial reasoning tasks. The team found that LLMs could handle 2D grid reasoning without explicit positional encodings, that symmetry augmentations served as a powerful validation mechanism rather than a training aid, and that the models second-best solutions were often conceptually correct just wrong in a specific detail like rotation direction. They also discuss why their fine-tuned 12B model outperformed much larger foundation models, the role of LoRA in preventing catastrophic forgetting during test-time training, and what the entropy distribution of their search trees reveals about how the model represents uncertainty across different task types.
REFERENCES:
Paper:
[00:01:00] The ARChitects: Winning ARC Prize 2024 Solution
https://github.com/da-fr/arc-prize-2024/blob/main/the_architects.pdf
[00:03:38] Robustness of Analogical Reasoning in LLMs
https://arxiv.org/html/2411.14215
[00:14:58] Search Methods in Language Models
https://arxiv.org/html/2408.00724v2
[00:22:28] GPT-4 Code Solutions for ARC (50% SOTA)
https://redwoodresearch.substack.com/p/getting-50-sota-on-arc-agi-with-gpt
[00:53:08] Overcoming Catastrophic Forgetting
https://www.pnas.org/doi/10.1073/pnas.1611835114
[00:53:58] LoRA: Low-Rank Adaptation of Large Language Models
https://arxiv.org/abs/2106.09685
Tool:
[00:07:48] Re-ARC Dataset Generator
https://github.com/michaelhodel/re-arc
LINKS:
Full Transcript: https://app.rescript.info/share/57e5d773f2d0b195cbce7eee1f53aef2
Download PDF transcript: https://app.rescript.info/api/public/sessions/7772acbf1f11f44b/pdf
Daniel Franzen
https://github.com/da-fr
REFS
[00:01:05] Winning ARC 2024 solution using 12B param model, Franzen, Disselhoff, Hartmann
https://github.com/da-fr/arc-prize-2024/blob/main/the_architects.pdf
[00:07:50] Re-ARC dataset generator for ARC task variations, Michael Hodel
https://github.com/michaelhodel/re-arc
[00:22:30] GPT-4 guided code solutions for ARC tasks, Ryan Greenblatt
https://redwoodresearch.substack.com/p/getting-50-sota-on-arc-agi-with-gpt")
