Conformer2
Meet Conformer-2: Advanced Speech Recognition Featuring Greater Accuracy and Faster Processing
About
Presenting Conformer-2, our newest AI model for automatic speech recognition. Trained on 1.1M hours of English audio data, it extends Conformer-1 with enhancements in proper nouns, alphanumerics, and noise robustness. Conformer-2 advances our original Conformer-1 release by boosting both model performance and speed. This update delivers a 31.7% improvement on alphanumerics, a 6.8% improvement on Proper Noun Error Rate, and a 12.0% improvement in robustness to noise. These gains result from expanding training data to 1.1M hours and employing more models for pseudo-labeling data. Conformer-1 set state-of-the-art performance with strong noise robustness, ideal for real-world audio conditions. Conformer-2 matches Conformer-1's word error rate while advancing user-oriented metrics. Since Conformer-1's release, our engineering team has reduced inference pipeline latency by up to 53.7%.
Details
Conformer-2 is a state-of-the-art speech recognition model developed by AssemblyAI, trained on 1.1 million hours of English audio data. It builds on Conformer-1 with significant improvements in proper nouns (6.8% error rate reduction), alphanumerics (31.7% improvement), and noise robustness (12.0% improvement), while maintaining word error rate parity. The model leverages model ensembling for pseudo-labeling and scaling laws inspired by DeepMind's Chinchilla paper. Additionally, inference latency has been reduced by up to 53.7% compared to Conformer-1, making it suitable for real-time applications. Conformer-2 is designed for real-world audio conditions and is available through AssemblyAI's API.
Key Features
- 31.7% improvement on alphanumerics
- 6.8% improvement on Proper Noun Error Rate
- 12.0% boost in noise robustness
- Trained on 1.1M hours of English audio
- Maintains word error rate parity with Conformer-1
- Up to 53.7% reduction in latency
- Enhanced performance in real-world audio conditions
- Improved transcription accuracy
- Increased number of models used for pseudo-labeling data
- Developed by AssemblyAI
Use Cases
- Podcasters: Enhance transcription accuracy for podcasts, especially with proper nouns and alphanumerics.
- Business professionals: Improve the transcription quality of business meetings and conferences with enhanced noise robustness.
- Media creators: Accurately transcribe video content, ensuring proper noun correctness and alphanumeric accuracy.
- Researchers: Transcribe interviews and focus group discussions with high accuracy, even in noisy environments.
- Customer service teams: Utilize accurate transcriptions for customer interactions, improving service quality and response time.
- Educational institutions: Provide precise transcriptions for online courses and lectures, accommodating proper nouns and complex alphanumerics.
- Journalists: Obtain reliable transcriptions of interviews and reports, maintaining the integrity of proper nouns.
- Healthcare providers: Accurately transcribe medical dictations and notes, supporting critical alphanumeric and proper noun recognition.
- Market researchers: Analyze consumer feedback and focus group recordings with high transcription accuracy in various environments.
- Software developers: Integrate advanced speech recognition into applications, delivering accurate transcriptions in real-time scenarios.
Pros & Cons
Pros
- State-of-the-art speech recognition performance
- Significant improvements on proper nouns and alphanumerics
- Excellent noise robustness for real-world conditions
- Reduced inference latency (up to 53.7%)
- Uses advanced model ensembling for better accuracy
- Maintains high accuracy while improving user-oriented metrics
Cons
- Proprietary model only available via AssemblyAI API
- Requires internet connection for API usage
- Pricing requires contacting sales (no self-serve tier listed)
Frequently Asked Questions
What is Conformer-2?
Conformer-2 is AssemblyAI's latest AI model for automatic speech recognition, designed to improve performance on proper nouns, alphanumerics, and noise robustness.
How much training data was used for Conformer-2?
Conformer-2 was trained on 1.1 million hours of English audio data.
What improvements does Conformer-2 offer over Conformer-1?
Conformer-2 offers a 31.7% improvement on alphanumerics, a 6.8% improvement on Proper Noun Error Rate, a 12.0% boost in noise robustness, and a significant reduction in latency.
Does Conformer-2 maintain the word error rate of Conformer-1?
Yes, Conformer-2 maintains parity with Conformer-1 in terms of word error rate while providing additional improvements.
What are the key features of Conformer-2?
Key features of Conformer-2 include enhanced performance on alphanumerics, better recognition of proper nouns, improved noise robustness, and reduced latency.
Who developed Conformer-2?
Conformer-2 was developed by AssemblyAI, an industry leader in AI models for automatic speech recognition.
How does Conformer-2 handle noisy environments?
Conformer-2 is designed to be more robust to noise, with a 12.0% improvement in noise robustness over Conformer-1.
What is the latency improvement in Conformer-2?
The latency of the inference pipeline has been reduced by up to 53.7% since the release of Conformer-1.
Is Conformer-2 suitable for real-world audio conditions?
Yes, Conformer-2 is designed to be highly robust and suitable for real-world audio conditions, maintaining strong noise robustness.
What is the Proper Noun Error Rate improvement in Conformer-2?
Conformer-2 achieves a 6.8% improvement in Proper Noun Error Rate over its predecessor, Conformer-1.