Voice Surge: Drama 3, xAI TTS, Sarvam Bulbul V4 Flash, HeyGen Voice, SGLang-Omni and Low-Cost Streaming Voice Stacks
Key Questions
Which TTS and STT models currently lead in rankings?
Cartesia Sonic 3.5 and Ink 2 are ranked #1 for streaming TTS and STT respectively. Their pricing has not yet been disclosed.
What new voice features are coming from ElevenLabs?
ElevenLabs announced Dubbing v2 and Eleven v4 TTS preview at the Warsaw Summit 2026. Details on availability and pricing were not specified in the summary.
How does Cohere's new ASR model perform?
Cohere Transcribe is an open-source 2B ASR model that ranks #1 on the Hugging Face Far-Field benchmark.
When will NLPearl Speech-to-Speech v2 launch?
NLPearl Speech-to-Speech v2 is scheduled to launch on August 1st with claims of human-like emotion and 800ms latency.
What resources are available for voice agent infrastructure?
Modal published a deep-dive on voice agent infrastructure with actionable latency and cost patterns for developers.
语音 API 进入“情绪表达、区域语言和实时流式”并行竞争阶段。Drama 3 开放权重及约 80% API 折扣、xAI TTS 约 $0.015/1K 字符、Sarvam Bulbul V4 Flash 约 ₹20/万字符，为个人团队提供清晰的低成本产品入口；HeyGen Voice 则提供 39 语言但延迟约 6.6 秒，适合高价值视频配音。SGLang-Omni 和 Gradium 的案例进一步强化了多供应商路由、TTFA 优化和自建基础设施的重要性。