May 2026 Summaries

Phonon, a 100M-parameter on-device Text-To-Speech (TTS) model, has achieved a significant milestone by reaching a 1.00% word error rate (WER) on the Seed-TTS English benchmark, outperforming larger models like NeuTTS Air (552M), KaniTTS2 (450M), and NeuTTS Nano (229M). With voice cloning disabled and a fixed high-quality voice, Phonon further reduces its WER to 0.83%, surpassing models such as Kokoro and Magpie. The model is based on Continuous Audio Language Models with flow-matching for waveform generation and supports deployment scenarios unsuitable for cloud APIs, including offline and privacy-sensitive applications. Updates since April have seen improvements in WER, speaker similarity, and the removal of input length padding, alongside support for int8 quantization, enhancing inference speed without compromising audio quality. Phonon is in private beta, offering a robust edge-based solution for TTS tasks.

Gradium leads Coval's Text-to-Speech (TTS) benchmarks as of May 13, 2026, excelling in both latency and quality metrics, including Word Error Rate (WER). Coval, an independent evaluation platform founded by Brooke Hopkins in 2024, provides standardized and transparent TTS benchmarking across providers, contrasting with vendor-reported metrics that often lack comparability. Gradium's success is attributed to its innovative audio language models (ALMs) that integrate multiple voice tasks within a single architecture, originally pioneered by its founders at Google and Meta. Coval's benchmarks focus on Time to First Audio (TTFA), latency range, and WER, with Gradium's models excelling in these areas due to their Delayed Streams Modeling approach, which balances quality and latency in real-time TTS. The open-source nature of Coval's methodology allows for tailored evaluations, providing voice agent teams with relevant performance insights specific to their operational contexts.