Add a default 20% silence ratio to voice cost estimates

tools/cost_calculator/README.md

@@ -133,7 +133,7 @@ Global, Data Zone, Batch, or fine-tuning meters.
 ## Assumptions and formulas
 
 Defaults use the supplied **90-second call**, **4 user + 4 agent
-turns**, **45 seconds of speech each**, **100% reused-context caching**,
+turns**, **45 seconds of speech allocation each before silence**, **100% reused-context caching**,
 and **150 agent text tokens**
 for text-response pipelines. User transcript length was not supplied:
 the calculator explicitly assumes an editable **150 user text tokens**.
@@ -142,9 +142,25 @@ These defaults are modeling inputs, not measured usage.
 Audio defaults follow
 [Voice Live's Azure OpenAI estimates](https://learn.microsoft.com/en-us/azure/ai-services/speech-service/voice-live#token-usage-and-cost-estimation):
 **10 input audio tokens/sec** and **20 output audio tokens/sec**, giving
-**450 input** and **900 output** audio tokens per default call. Both rates
+**450 input** and **900 output** audio tokens before the silence adjustment. Both rates
 remain editable for other model families or measured usage.
 
+**Both-parties-silent ratio defaults to 20%** and is editable from 0 to 100%.
+The suggested 10-30% range is a planning assumption, not a universal average.
+Multiply each side's entered speech seconds by `1 - silencePercent / 100`
+before calculating audio tokens. The default therefore models **36 seconds
+per side**, **360 input tokens**, and **720 output tokens**, with 18 seconds
+of both-parties silence in the 90-second call.
+This scales fresh and replayed native audio, STT, and TTS usage only, not
+unit prices. Text, prompt, reasoning, turn count, GPT Live session minutes,
+and the full-call per-minute denominator remain unchanged. It is not a
+blanket cost discount or a guarantee that silence is unbilled. Set the ratio
+to **0%** when entered speech times already exclude silence to avoid double
+counting. With smaller baseline speech allocations, the ratio reduces only
+those allocations; it does not subtract additional time from the full call.
+The summary shows adjusted audio usage; CSV includes `silence_percent` and
+the complete assumptions.
+
 Each call starts cold; the system prompt costs `P * input_rate` on its
 first turn. Its `P * (turns - 1)` reused tokens are charged at the selected
 cache hit rate. With equal turn lengths, each side's total conversation

tools/cost_calculator/app.mjs

@@ -125,7 +125,7 @@ function renderOverview(a, model, pipeline, result, book) {
   badge.className = `badge${result.list === null ? "" : " good"}`;
   $("summary").textContent = result.missing.length ?
     `List estimate needs: ${result.missing.map(key => labels[key]).join(", ")}. Enter missing values in the price book below.` :
-    `${result.inputAudio.toLocaleString("en-US")} fresh caller audio tokens / ${result.outputAudio.toLocaleString("en-US")} generated speech tokens / ${a.turns} user + ${a.turns} agent turns.`;
+    `${result.inputAudio.toLocaleString("en-US")} fresh caller audio tokens / ${result.outputAudio.toLocaleString("en-US")} generated speech tokens / ${a.turns} user + ${a.turns} agent turns. Audio usage includes ${a.silencePercent}% silence reduction; text and session minutes are unchanged.`;
   $("model-note").textContent = pipeline === "live" ?
     "Backend model + GPT Live voice use Azure OpenAI South India Global references ($0.05/minute for voice). Availability must be confirmed." :
     book.usesGlobalReference ?
@@ -189,7 +189,7 @@ function renderComparison(a) {
         model: model.id, pipeline, hosting: effectiveHosting(pipeline), region: snapshot.region,
         pricing_basis: usesGlobalReference ? "aoai-global-model-plus-byo-speech-reference" :
           effectiveHosting(pipeline) === "managed" ? "voice-live-tier-reference" : "byom",
-        prompt_tokens: prompt, list_per_call: result.list,
+        prompt_tokens: prompt, silence_percent: a.silencePercent, list_per_call: result.list,
         list_per_minute: result.listPerMinute, missing_list_rates: result.missing.join("; "),
         price_snapshot: snapshot.retrievedAt,
         model_rate_region: modelMeter ? modelMeter.armRegionName || snapshot.region : "",

tools/cost_calculator/calculator.mjs

@@ -2,7 +2,7 @@ export const DEFAULTS = Object.freeze({
   duration: 90, turns: 4, humanSeconds: 45, agentSeconds: 45,
   inputAudioRate: 10, outputAudioRate: 20,
   userText: 150, agentText: 150, nativeText: 0, reasoning: 0,
-  cachePercent: 100, prompt: 5000, calls: 10000,
+  cachePercent: 100, silencePercent: 20, prompt: 5000, calls: 10000,
 });
 
 export const PROMPTS = [0, 1000, 5000, 10000, 20000];
@@ -39,6 +39,7 @@ export function validateAssumptions(a) {
     throw new Error("Human + agent speaking time cannot exceed call duration in this turn-based estimate.");
   }
   if (a.cachePercent > 100) throw new Error("Cache percentage must be between 0 and 100.");
+  if (a.silencePercent > 100) throw new Error("Silence percentage must be between 0 and 100.");
 }
 
 export function meterRate(snapshot, name, unit = "tokens") {
@@ -134,8 +135,9 @@ export function calculate(a, pipeline, rates) {
   const missing = new Set();
   const reuse = (a.turns - 1) / 2;
   const hit = a.cachePercent / 100;
-  const inputAudio = a.humanSeconds * a.inputAudioRate;
-  const outputAudio = a.agentSeconds * a.outputAudioRate;
+  const speakingFraction = 1 - a.silencePercent / 100;
+  const inputAudio = a.humanSeconds * speakingFraction * a.inputAudioRate;
+  const outputAudio = a.agentSeconds * speakingFraction * a.outputAudioRate;
   function add(label, quantity, key, unit = "tokens") {
     if (!Number.isFinite(quantity)) throw new Error(`Token quantity is too large: ${label}`);
     if (quantity === 0) return;

tools/cost_calculator/calculator.test.mjs

@@ -51,7 +51,7 @@ test("South India GPT-4.1 and realtime prices do not reuse managed tier rates",
   close(book("gpt-realtime-2.1-mini", "byom").textCached, 0.06);
   close(book("gpt-realtime", "byom").audioOutput, 64);
   const result = calculate({ ...DEFAULTS, prompt: 0 }, "cascade", rates);
-  close(result.list, (150 * 2 + 450 * 0.5 + 150 * 8 + 450 * 12.5 + 900 * 23) / 1e6);
+  close(result.list, (150 * 2 + 450 * 0.5 + 150 * 8 + 360 * 12.5 + 720 * 23) / 1e6);
 });
 
 test("unverified managed tiers use explicit Global model references, not invented tier rates", () => {
@@ -81,22 +81,23 @@ test("default native call counts fresh audio and cached history without speech d
   const result = calculate({ ...DEFAULTS, prompt: 0 }, "native", book("gpt-realtime-2.1"));
   assert.equal(DEFAULTS.inputAudioRate, 10);
   assert.equal(DEFAULTS.outputAudioRate, 20);
-  assert.equal(result.inputAudio, 450);
-  assert.equal(result.outputAudio, 900);
-  close(result.list, (450 * 32 + 900 * 64 + 2025 * 0.4) / 1e6);
+  assert.equal(DEFAULTS.silencePercent, 20);
+  assert.equal(result.inputAudio, 360);
+  assert.equal(result.outputAudio, 720);
+  close(result.list, (360 * 32 + 720 * 64 + 1620 * 0.4) / 1e6);
   assert.ok(result.rows.every(r => !r.label.includes("Azure")));
 });
 
 test("input audio rate overrides do not change output audio usage", () => {
   const rates = book("gpt-realtime-2.1");
   const original = calculate({ ...DEFAULTS, prompt: 0, inputAudioRate: 20 }, "native", rates);
   const updated = calculate({ ...DEFAULTS, prompt: 0 }, "native", rates);
-  assert.equal(original.inputAudio, 900);
-  assert.equal(updated.inputAudio, 450);
-  assert.equal(original.outputAudio, 900);
-  assert.equal(updated.outputAudio, 900);
-  close(original.list, (900 * 32 + 900 * 64 + 2700 * 0.4) / 1e6);
-  close(original.list - updated.list, (450 * 32 + 675 * 0.4) / 1e6);
+  assert.equal(original.inputAudio, 720);
+  assert.equal(updated.inputAudio, 360);
+  assert.equal(original.outputAudio, 720);
+  assert.equal(updated.outputAudio, 720);
+  close(original.list, (720 * 32 + 720 * 64 + 2160 * 0.4) / 1e6);
+  close(original.list - updated.list, (360 * 32 + 540 * 0.4) / 1e6);
 });
 
 test("system prompt is full price once, reused across the remaining turns", () => {
@@ -105,21 +106,21 @@ test("system prompt is full price once, reused across the remaining turns", () =
   const full = calculate(DEFAULTS, "native", rates);
   close(full.list - base.list, 5000 * (4 + 3 * 0.4) / 1e6);
   const noCache = calculate({ ...DEFAULTS, cachePercent: 0 }, "native", rates);
-  close(noCache.list, (20000 * 4 + 2475 * 32 + 900 * 64) / 1e6);
+  close(noCache.list, (20000 * 4 + 1980 * 32 + 720 * 64) / 1e6);
   const halfCache = calculate({ ...DEFAULTS, cachePercent: 50 }, "native", rates);
   close(halfCache.list, (noCache.list + full.list) / 2);
 });
 
 test("single turn does not charge replayed context", () => {
   const result = calculate({ ...DEFAULTS, turns: 1 }, "hybrid", book("gpt-realtime-2.1"));
   assert.ok(result.rows.every(r => !r.label.includes("/ cached")));
-  close(result.list, (5000 * 4 + 450 * 32 + 150 * 24 + 900 * 31) / 1e6);
+  close(result.list, (5000 * 4 + 360 * 32 + 150 * 24 + 720 * 31) / 1e6);
 });
 
 test("hybrid uses 150 text output tokens and HD list price, no native output audio charge", () => {
   const result = calculate({ ...DEFAULTS, prompt: 0 }, "hybrid", book("gpt-realtime-2.1"));
-  const llm = (450 * 32 + 675 * 0.4 + 225 * 0.4 + 150 * 24) / 1e6;
-  close(result.list, llm + 900 * 31 / 1e6);
+  const llm = (360 * 32 + 540 * 0.4 + 225 * 0.4 + 150 * 24) / 1e6;
+  close(result.list, llm + 720 * 31 / 1e6);
   assert.ok(result.rows.every(r => r.key !== "audioOutput"));
 });
 
@@ -130,7 +131,7 @@ test("BYOM uses BYO speech meters, not managed tier prices", () => {
   close(rates.textOutput, 15);
   const result = calculate({ ...DEFAULTS, prompt: 0 }, "cascade", rates);
   const llm = (150 * 2.5 + 450 * 0.25 + 150 * 15) / 1e6;
-  close(result.list, llm + (450 * 12.5 + 900 * 23) / 1e6);
+  close(result.list, llm + (360 * 12.5 + 720 * 23) / 1e6);
   close(result.listPerMinute, result.list / 1.5);
 });
 
@@ -204,7 +205,8 @@ test("results and components expose only list-price fields", () => {
 
 test("invalid assumptions and negative rates are rejected", () => {
   for (const changes of [{ duration: 0 }, { turns: 0 }, { turns: 1.5 }, { humanSeconds: 91 },
-    { agentSeconds: -1 }, { prompt: NaN }, { cachePercent: 101 }, { calls: 1.5 }, { inputAudioRate: Infinity }]) {
+    { agentSeconds: -1 }, { prompt: NaN }, { cachePercent: 101 }, { calls: 1.5 }, { inputAudioRate: Infinity },
+    { silencePercent: -1 }, { silencePercent: 101 }, { silencePercent: NaN }, { silencePercent: Infinity }]) {
     assert.throws(() => calculate({ ...DEFAULTS, ...changes }, "native", book("gpt-realtime")));
   }
   assert.throws(() => calculate(DEFAULTS, "native", { ...book("gpt-realtime"), audioInput: -1 }), /Invalid rate/);
@@ -215,7 +217,40 @@ test("zero-priced input is explicit and unused missing rates do not block a tota
   const rates = { ...book("gpt-realtime"), textInput: null, textCached: null, audioOutput: 0 };
   const result = calculate({ ...DEFAULTS, prompt: 0 }, "native", rates);
   assert.deepEqual(result.missing, []);
-  close(result.list, (450 * 32 + 2025 * 0.4) / 1e6);
+  close(result.list, (360 * 32 + 1620 * 0.4) / 1e6);
+});
+
+test("silence scales only audio rows, with zero preserving baseline and full silence removing audio", () => {
+  const audioKeys = new Set(["audioInput", "audioCached", "audioOutput", "speechInput", "speechOutput"]);
+  for (const hosting of ["managed", "byom"]) {
+    for (const [id, pipeline] of [["gpt-realtime-2.1", "native"], ["gpt-realtime-2.1", "hybrid"],
+      ["gpt-4.1", "cascade"], ["gpt-6-sol", "live"]]) {
+      const rates = book(id, hosting, pipeline);
+      const a = { ...DEFAULTS, nativeText: 100, reasoning: 80 };
+      const baseline = calculate({ ...a, silencePercent: 0 }, pipeline, rates);
+      assert.equal(baseline.inputAudio, 450);
+      assert.equal(baseline.outputAudio, 900);
+      for (const silencePercent of [10, 20, 30, 100]) {
+        const result = calculate({ ...a, silencePercent }, pipeline, rates);
+        const fraction = 1 - silencePercent / 100;
+        close(result.inputAudio, 450 * fraction);
+        close(result.outputAudio, 900 * fraction);
+        for (const row of baseline.rows) {
+          const actual = result.rows.find(r => r.label === row.label);
+          const factor = audioKeys.has(row.key) ? fraction : 1;
+          close(actual?.quantity ?? 0, row.quantity * factor);
+          close(actual?.list ?? 0, row.list * factor);
+          if (actual) assert.equal(actual.rate, row.rate);
+        }
+        const audioCost = baseline.rows.filter(r => audioKeys.has(r.key)).reduce((sum, row) => sum + row.list, 0);
+        close(result.list, baseline.list - audioCost * silencePercent / 100);
+        close(result.listPerMinute, result.list / 1.5);
+        if (pipeline === "live") close(result.list, baseline.list);
+      }
+    }
+  }
+  close(calculate({ ...DEFAULTS, prompt: 0, silencePercent: 0 }, "native", book("gpt-realtime-2.1")).list,
+    (450 * 32 + 900 * 64 + 2025 * 0.4) / 1e6);
 });
 
 test("all five prompt sizes grow monotonically for every priced model", () => {

tools/cost_calculator/index.html

@@ -29,8 +29,9 @@ <h2><span>02</span> Conversation assumptions</h2>
           <div class="fields">
             <label>Call duration (sec)<input name="duration" type="number" min="0.01" step="any" required></label>
             <label>User / agent turn pairs<input name="turns" type="number" min="1" step="1" required></label>
-            <label>Human speech (sec)<input name="humanSeconds" type="number" min="0" step="any" required></label>
-            <label>Agent speech (sec)<input name="agentSeconds" type="number" min="0" step="any" required></label>
+            <label>Human speech before silence (sec)<input name="humanSeconds" type="number" min="0" step="any" required></label>
+            <label>Agent speech before silence (sec)<input name="agentSeconds" type="number" min="0" step="any" required></label>
+            <label>Both parties silent (%)<input name="silencePercent" type="number" min="0" max="100" step="any" required></label>
             <label>Input audio tokens/sec<input name="inputAudioRate" type="number" min="0" step="any" required></label>
             <label>Output audio tokens/sec<input name="outputAudioRate" type="number" min="0" step="any" required></label>
             <label>User text tokens / call<input name="userText" type="number" min="0" step="any" required></label>
@@ -40,7 +41,7 @@ <h2><span>02</span> Conversation assumptions</h2>
             <label>System prompt tokens<input name="prompt" type="number" min="0" step="any" required></label>
             <label>Context cache hit (%)<input name="cachePercent" type="number" min="0" max="100" step="any" required></label>
           </div>
-          <p class="hint">Defaults: 10 input and 20 output audio tokens/sec, giving 450 input + 900 output audio tokens. Text totals are independent estimates, not derived from audio. The 150-token user transcript is illustrative.</p>
+          <p class="hint">Default 20% silence reduces 45 sec per side to 36 sec: 360 input + 720 output audio tokens at 10 input / 20 output tokens/sec. A 10-30% silence range is a planning assumption, not a universal average. Set silence to 0% if your speech times already exclude silence. Text totals and GPT Live session minutes are not discounted.</p>
         </section>
         <section>
           <h2><span>03</span> Call volume</h2>
@@ -94,6 +95,7 @@ <h2><span>03</span> Call volume</h2>
         <h2>Read the estimate, not an invoice</h2>
         <p>Each call starts with an uncached prompt. On later turns the prompt and prior conversation are reused at your cache-hit percentage. With equally sized turns, prior caller/agent tokens are each replayed (turns - 1) / 2 times. Cache misses use full input rates.</p>
         <p>Native mode reuses audio history; text-response mode reuses caller audio and agent text; cascade and reasoning backends reuse text. Only newly generated speech is synthesized. Reasoning tokens are charged as text output, not retained as visible conversation history.</p>
+        <p>Silence is a usage assumption, not a price discount or billing guarantee. Each side's entered speech seconds is multiplied by (1 - silence % / 100) before computing fresh and replayed audio tokens, STT, and TTS. Text, prompt, reasoning, turn count, GPT Live session minutes, and the full-call per-minute denominator stay unchanged. With the default 90-second speaking allocation, 20% represents 18 seconds with both parties silent; avoid counting silence twice when entering measured speech times.</p>
         <p>100% caching is an idealized assumption, not guaranteed prefix-cache eligibility. Audio defaults use the public guide's estimates of 10 input and 20 output audio tokens/sec for Azure OpenAI models. Actual tokens, silence billing, tools, storage, telephony, hosting capacity, tax, and network charges can change the result and are excluded.</p>
         <p>Model selection is for economic modeling, not a guarantee of South India deployment availability or supported BYOM audio connections. South India Global is the reference for BYOM models and GPT Live. No South India Voice Live managed-tier or BYO speech meters were found, so those use approved Central India references. Azure HD speech output also deliberately uses Central India rates. GPT-6 is explicitly GPT-6 Sol, short-context Standard pricing. GPT Live voice uses the South India Global rate ($3/hour = $0.05/minute); backend model usage is additional and no Azure HD charge is added. GPT Live's delegation frequency and backend context differ in practice; this estimate uses the same turn-based backend assumptions and a session-minute rate.</p>
         <p>If a model has no verified managed tier, its economic reference combines AOAI Global model rates with BYO speech rates and is labeled accordingly. It is not a managed Voice Live price or support claim.</p>