Publish voice duplex architecture latency training and evaluation research

model.js

@@ -0,0 +1,61 @@
+export function queueSteps(chunkMs, computeMs, count = 8) {
+  if (
+    !Number.isFinite(chunkMs) ||
+    !Number.isFinite(computeMs) ||
+    chunkMs <= 0 ||
+    computeMs <= 0
+  ) {
+    throw new RangeError("Time values must be finite and positive.");
+  }
+  let previousEnd = 0;
+  return Array.from({ length: count }, (_, index) => {
+    const arrival = (index + 1) * chunkMs;
+    const start = Math.max(arrival, previousEnd);
+    const end = start + computeMs;
+    previousEnd = end;
+    return { index: index + 1, arrival, start, end, wait: start - arrival };
+  });
+}
+
+export const architectures = {
+  cascade: {
+    nodes: [
+      ["用户语音", "声音输入", "input"],
+      ["ASR", "识别成文本"],
+      ["文本 LLM", "生成回答"],
+      ["TTS", "合成声音"],
+      ["系统语音", "播放输出", "output"],
+    ],
+    detail:
+      "<p><strong>两个独立的文本接口。</strong>ASR 把用户语音交给文本模型，TTS 再把回答读出。模块可以独立替换。仅传文字时，输入声学信息会在接口处被压缩。</p><p>持续双工需要在这些模块之间保持新输入、话轮状态、取消与播放进度的一致。</p>",
+  },
+  two: {
+    nodes: [
+      ["用户语音", "声音输入", "input"],
+      ["语音理解 LLM", "直接读取音频特征"],
+      ["回答文本", "内容接口"],
+      ["TTS", "合成声音"],
+      ["系统语音", "播放输出", "output"],
+    ],
+    detail:
+      "<p><strong>输入端直接理解声音。</strong>模型接收音频特征并生成回答文字，之后由独立 TTS 合成。输入端可以保留语气等线索，输出端仍通过文本接口组织。</p><p>本页以 Ultravox 核心加外接 TTS 作为这一工作定义的例子。</p>",
+  },
+  native: {
+    nodes: [
+      ["用户语音", "声音输入", "input"],
+      ["语音表示", "编码或特征"],
+      ["联合交互模型", "声音与内容建模"],
+      ["语音解码", "还原波形"],
+      ["系统语音", "播放输出", "output"],
+    ],
+    detail:
+      "<p><strong>语音直接参与内容与声学生成。</strong>编码器、音频 token、文字辅助流和波形解码器可以共同存在，具体组织依模型而异。</p><p>持续听说依赖输入更新方式、因果训练与实时服务。原生语音接口与全双工行为分别核验。</p>",
+  },
+};
+
+export const benchmarks = [
+  { name: "级联管线", first: 8.78, answer: 10.12 },
+  { name: "Ultravox 0.7", first: 3.88, answer: 8.4 },
+  { name: "GPT-Realtime 1.5", first: 6.36, answer: 6.89 },
+  { name: "Gemini Live 3.1", first: 3.95, answer: 4.25 },
+];