📝 Blogs & Articles — Awesome Speech Audio
Loading posts…
Loading posts…
14 posts, articles, and resources from across the field.
The field is focusing on improving automatic speech recognition systems through benchmarking and real-world applications.
3 postsResearch is exploring enhancements in the quality and expressiveness of text-to-speech systems.
2 postsThere is a growing interest in live music models and APIs for real-time audio generation.
3 postsWhy it matters — This post highlights the challenges of training ASR models for clinical terminology, emphasizing the need for specialized techniques to enhance recognition accuracy in healthcare settings.
Why it matters — The use of low-rank adaptation and data augmentation in LLM-based TTS systems addresses the need for high-quality, accent-free speech synthesis across multiple languages, which is crucial for global applications.
Why it matters — Gemini 3.1's granular audio tags represent a significant advancement in controlling expressiveness in AI-generated speech, allowing for more nuanced audio outputs that can enhance user experience.
Why it matters — The FFASR Leaderboard provides a standardized framework for evaluating ASR systems in real-world scenarios, which is essential for researchers to benchmark their models against industry standards.
Why it matters — Magenta RealTime 2 introduces a powerful tool for creating AI musical instruments, which can inspire new research directions in real-time music generation and interaction.
Why it matters — The addition of Benchmaxxer Repellant to the Open ASR Leaderboard highlights advancements in ASR performance, offering researchers insights into effective techniques for improving recognition accuracy.
Why it matters — Gemini's new music generation capabilities expand the creative potential of AI, allowing researchers to explore the intersection of text, images, and music generation.
Why it matters — Lyria Camera's integration of music generation with image understanding presents innovative opportunities for researchers to investigate multimodal AI applications.
Why it matters — Magenta RealTime's open-weights model facilitates interactive music creation, providing a platform for researchers to experiment with live performance and AI collaboration.
Why it matters — The Lyria RealTime API opens new avenues for researchers to explore generative music technologies, enhancing the creative possibilities in music generation.
Why it matters — The introduction of StruQ and SecAlign addresses the critical issue of prompt injection attacks in LLMs, providing researchers with strategies to improve the security of LLM-integrated applications.
Why it matters — Dan Deacon's collaboration with MusicLM showcases the practical applications of AI in music composition, offering insights into how AI can enhance artistic processes.
Language models play a central role in automatic speech recognition (ASR), yet most methods rely on text-only models unaware of ASR error patterns. Recently, large language models (LLMs) have been applied to ASR correction, but introduce latency and hallucination concerns. We revisit ASR error correction with compact seq2seq models, trained on ASR errors from real and synthetic audio. To scale training, we construct synthetic corpora via cascaded TTS and ASR, finding that matching the diversity of realistic error distributions is key. We propose correction-first decoding, where the correction…