March 2023 Summaries

The article presents a comprehensive guide to accurately timing individual operations in a computational graph, particularly for machine learning models on GPUs. It highlights the importance of host-device synchronization, CUDA events, warm-up steps, fixed clocks, cache flush, and sleep/CUDA graphs in achieving accurate and repeatable results. The guide provides examples and tips specific to PyTorch, but the principles discussed can be applied to CUDA programming in general.

Ursa is Speechmatics' latest speech-to-text system that can transcribe difficult audio with incredible accuracy regardless of demographics, which is crucial for high-quality downstream performance. Large language models like ChatGPT and GPT4 are trained to predict the next word given the sequence of words that have come before, learning from vast amounts of training data to perform tasks such as summarization, sentiment analysis, emotion detection, named entity recognition, and question answering. However, these models can gloss over some recognition errors and produce "better than input" answers due to hallucinations based on their knowledge from training data. The accuracy of the ASR transcript is crucial for ensuring a high-quality output, with Ursa producing transcripts with excellent accuracy particularly on named entities, technical terminology, and difficult audio. In contrast, lower-accuracy transcripts can cause errors ranging from spelling mistakes to complete inability to perform tasks, as demonstrated by experiments using GPT4 and ChatGPT on Ursa and Google transcriptions.

March 14 was a significant day for the AI community, with OpenAI releasing GPT-4, a multi-modal language model that combines text and image capabilities. GPT-4 has a longer context length than its predecessor, allowing it to process hundreds of pages in a single prompt. The model demonstrates impressive behavior such as visual question answering and image captioning. Its training dataset is likely similar to that of KOSMOS-1, which also uses pre-trained image encoders and multimodal inputs. GPT-4 has been fine-tuned using Reinforcement Learning from Human Feedback (RLHF) to align its output with user intent. The model's capabilities include text-only and text-vision tasks, with the latter demonstrating impressive visual reasoning abilities. As future models are trained on additional modalities like audio and video, they may possess even more advanced capabilities, such as generating art or music from text prompts. However, this growth in capabilities also raises concerns about AI safety, particularly "intent alignment," which requires careful consideration to ensure that systems optimize for intended goals rather than unintended ones. Overall, GPT-4 represents a significant step forward in multi-modal language modeling and highlights the need for continued research into AI safety and ethics.

Our first release of the year is focused on improving speech recognition accuracy, inclusivity, and reducing bias with new features such as GPU support, Ursa generation models, and a new Translation offering that integrates translation capabilities into our single Speech API. This update includes improvements to numeral formatting for English, automatic language identification, speaker diarization, and more, aiming to unlock human potential and increase accessibility for businesses in global markets. With the release of Ursa generation models, we've achieved significant accuracy uplifts, with an average 22% relative improvement for the Enhanced model and 35% relative improvement for the Standard model, using GPUs to enable larger machine learning models in production.

Ursa`, a speech-to-text technology, has achieved unprecedented accuracy with a 22% lead over its nearest competitor, breaking accessibility barriers in speech technologies. Its performance is notably accurate across various demographics, including age, gender, skin tone, socio-economic status, and level of education. Ursa's self-supervised learning approach enables it to learn from diverse audio data, significantly reducing bias in speech recognition. The technology demonstrates significant advancements in understanding every voice, with a focus on promoting inclusivity and accessibility for all users.

Ursa provides the world's most accurate speech-to-text technology, achieving human-level transcription accuracy on the Kincaid46 dataset and surpassing Microsoft by 22% in relative accuracy improvement. The system uses a self-supervised learning approach with GPUs for inference, scaling up models to 2 billion parameters, which significantly improves accuracy and reduces training time. Ursa's enhanced model outperforms other vendors, including Google, Amazon, and Whisper, with a 35% relative improvement over the previous release. Additionally, the system offers translation capabilities between English and 34 languages, achieving higher BLEU scores than Google in downstream tasks. Ursa represents a quantum leap forward in speech technologies, setting a new standard for the industry, and is the clear choice for anyone seeking best-in-class speech recognition and translation.