July 2022 Summaries

This week's recaps cover two Deep Learning papers: MAESTRO and Masked Autoencoders that Listen. The first paper proposes a method for learning unified representations from speech and text modalities, outperforming the current State-of-the-Art in ASR tasks. Key findings include the incorporation of lexical information using text-only inputs, improved performance in monolingual and multilingual setups, and efficient representation unification with minimal supervised data. In the second paper, a novel extension of masked autoencoders to audio is presented. The model works by splitting mel spectrograms into patches, masking most patches, and reconstructing them using an encoder-decoder approach. Key findings include the possibility of extending this method to temporal information like audio and video, extremely high patching ratios leading to more robust models in quality and bias settings, and local attention outperforming global for speech domains.

G2, the online software marketplace, has recognized AssemblyAI as a Summer 2022 High Performer and Momentum Leader in Voice Recognition Software category. This marks the fourth recognition by G2 for AssemblyAI, which is known for its transcription accuracy, support quality, continuous feature updates, and commitment to Deep Learning research. The company's Speech-to-Text API automatically converts audio and video content to highly accurate text, providing high ROI features to users.

1. Accurate Transcription: Hiring Intelligence Platforms must use accurate, reliable Speech-to-Text technology to create readable transcripts of interviews. Start-of-the-art Speech Recognition APIs like Automatic Casing and Punctuation models, Paragraph Detection features, and Speaker Diarization APIs increase the readability and accuracy of these transcriptions, making it easier for hiring managers to review candidate responses. 2. Generate Highlights and Key Analysis: In addition to providing an accurate, highly readable transcript, Hiring Intelligence Platforms must generate highlights and key analysis for users. This could include automatically searching transcripts for relevant skills or experience, creating highlight reels of key talking points, or analyzing a candidate’s overall behavior to determine best fit for the available role. Audio Intelligence APIs like Auto Chapters/Text Summarization APIs, Entity Detection APIs, Topic Detection APIs, and Sentiment Analysis APIs can be used to identify and label important information in interview transcripts, helping hiring managers more easily review candidate responses and make smarter hiring decisions. 3. Categorize, Tag, and Search Insights: Finally, Hiring Intelligence Platforms can use Speech Transcription and Audio Intelligence to categorize, tag, and search candidate interviews. For example, platforms could use this data to generate indexable categories that users can then use to tag or search, similar to how users use hashtags on platforms like Twitter. These “auto” or “smart” tags can be used for internal collaboration or for automatically attaching notes to indexed sections of interviews. Smart tags also make searching through and screening responses much faster and more efficient for hiring managers.

AssemblyAI has recently announced a $30M Series B funding round, led by global software investor Insight Partners, bringing their total funds raised to over $65M. This investment will support the company's efforts in leveraging AI technology for transcribing, understanding, and analyzing audio and video data. AssemblyAI has seen significant growth with its APIs used by developers and product teams across various industries. The company plans to use the new funds to accelerate its product roadmap, expand its AI infrastructure, and grow its research team. With over 30 open roles, AssemblyAI is seeking talented professionals to join their rapidly growing team.

The paper "Prune Once For All: Sparse Pre-Trained Language Models" introduces an architecture-agnostic method of training sparse pre-trained language models, allowing for pruning only during the pre-training phase. This technique results in better compression-to-accuracy ratios and eliminates the need to reconsider the model's architecture or task when applying pruning techniques during fine-tuning. The best scores were achieved with 85% and 90% weight pruning, while Quantized Aware Training (QAT) with 85% pruning led to an even more accurate and smaller model.