Product Overview: Conformer-2
Introduction
Conformer-2 is the latest advancement in automatic speech recognition technology developed by AssemblyAI. Building on the success of its predecessor, Conformer-1, this model is designed to significantly enhance performance in various critical areas, making it a state-of-the-art solution for speech-to-text applications.
What Conformer-2 Does
Conformer-2 is an AI model specifically engineered to improve the accuracy and efficiency of speech recognition. It is trained on an extensive dataset of 1.1 million hours of English audio, which is a substantial increase from the data used for Conformer-1. This extensive training enables the model to handle real-world audio conditions with unprecedented accuracy and robustness.
Key Features and Functionality
Improved Recognition Accuracy
- Alphanumerics Recognition: Conformer-2 achieves a 31.7% improvement in recognizing alphanumerics, which is crucial for applications requiring precise transcription of numbers, codes, and other alphanumeric data.
- Proper Noun Recognition: The model shows a 6.8% improvement in the recognition of proper nouns, significantly enhancing the accuracy of transcribing names, places, and other specific terms.
Enhanced Noise Robustness
- Conformer-2 boasts a 12.0% boost in noise robustness, allowing it to perform better in noisy environments and real-world audio conditions.
Speed and Latency
- Despite the increased model size and complexity, Conformer-2 reduces latency by up to 53.7%, making it faster and more efficient compared to its predecessor.
Model Size and Training
- The model has been expanded to 450 million parameters and trained on a massive 1.1 million hours of English audio data. This extensive training, combined with noisy student-teacher training methods, enhances the model’s performance across various domains and benchmarks.
Cost Control and Efficiency
- Conformer-2 introduces a new parameter called “Speech thresholds,” which allows users to control transcription costs by setting a minimum number of minutes required before processing a file. This feature is particularly useful for managing costs when dealing with sleep podcasts, music, or empty audio files.
User-Friendly Integration
- The model is already the default speech recognition model on Assembly AI’s API, making it easy for users to integrate and start using its advanced capabilities immediately. The Assembly AI playground and comprehensive documentation also facilitate a smooth onboarding process.
Benefits and Applications
- Improved Transcription Accuracy: Conformer-2’s enhanced performance in recognizing alphanumerics and proper nouns, along with its improved noise robustness, makes it highly effective for a wide range of applications, including virtual meeting transcription, podcast analysis, and other speech-to-text needs.
- Cost Efficiency: The ability to control transcription costs through Speech thresholds helps users optimize their expenses while maintaining high-quality transcription.
- Scalability and Performance: The model’s reduced latency and improved performance make it suitable for large-scale applications and real-time speech recognition tasks.
In summary, Conformer-2 represents a significant advancement in speech recognition technology, offering improved accuracy, speed, and cost efficiency, making it an indispensable tool for any application requiring high-quality speech-to-text transcription.