Voice Surge: Drama 3, xAI TTS, Sarvam, HeyGen and Kotoba
Key Questions
What new TTS models are highlighted in the voice surge?
Key releases include Lightning V3.1 Pro, Noiz AI with 3-second cloning at 80% lower cost than ElevenLabs, xAI Grok TTS, VoxCPM2, VibeVoice, Supertonic 3 supporting 31 languages, and NVIDIA's lightweight Kokoro TTS at 82M parameters.
How does StepAudio 2.5 perform on benchmarks?
StepAudio 2.5 is a unified ASR/TTS/live dialogue model trending on Hugging Face with a 3% CER and 68% win rate, positioning it as a strong option for low-cost voice SaaS wrappers.
What infrastructure tips are provided for voice agents?
Modal's guide covers latency optimizations such as GPU snapshotting to reduce cold starts from 25-90s to under 10s, warm CPU buffering, and cost patterns like batch sorting and isolated scaling for voice agent deployments.
What is the status of Zonos TTS updates?
Zonos 2, an update to Zyphra's TTS, is now available on Hugging Face Spaces with voice cloning and emotion sliders, expanding accessible options for indie voice wrappers.
Which open-source ASR model leads recent benchmarks?
Cohere Transcribe, a 2B parameter open-source model, ranks #1 on the new Hugging Face Far-Field ASR benchmark and supports 14 languages as a low-cost building block for voice applications.
语音 API 竞争仍集中于表现力、区域语言、流式体验和单位成本。Grok Voice 3.0 的 $0.08/分钟信号伴随首音延迟略升和 latest 别名切换风险；新近的架构比较也提示 speech-to-speech 与级联方案在延迟、计费和转录控制上各有取舍，供应商数字需核实。