brianyin/ajs-298-unify-all-metric-duration-to-ms (#765)

toubatbrian · web-flow · commit 5358bb8308eb · 2025-10-12T23:44:46.000-07:00
diff --git a/.changeset/hungry-olives-sink.md b/.changeset/hungry-olives-sink.md
@@ -0,0 +1,7 @@
+---
+'@livekit/agents-plugin-google': patch
+'@livekit/agents-plugin-openai': patch
+'@livekit/agents': patch
+---
+
+Convert and rename all time-based metric fields to \*Ms variants
diff --git a/agents/src/audio.ts b/agents/src/audio.ts
@@ -5,7 +5,7 @@ import { AudioFrame } from '@livekit/rtc-node';
 import { log } from './log.js';
 import type { AudioBuffer } from './utils.js';
 
-export function calculateAudioDuration(frame: AudioBuffer) {
+export function calculateAudioDurationSeconds(frame: AudioBuffer) {
   // TODO(AJS-102): use frame.durationMs once available in rtc-node
   return Array.isArray(frame)
     ? frame.reduce((sum, a) => sum + a.samplesPerChannel / a.sampleRate, 0)
diff --git a/agents/src/llm/llm.ts b/agents/src/llm/llm.ts
@@ -203,21 +203,24 @@ export abstract class LLMStream implements AsyncIterableIterator<ChatChunk> {
     this.output.close();
 
     const duration = process.hrtime.bigint() - startTime;
+    const durationMs = Math.trunc(Number(duration / BigInt(1000000)));
     const metrics: LLMMetrics = {
       type: 'llm_metrics',
       timestamp: Date.now(),
       requestId,
-      ttft: ttft === BigInt(-1) ? -1 : Math.trunc(Number(ttft / BigInt(1000000))),
-      duration: Math.trunc(Number(duration / BigInt(1000000))),
+      ttftMs: ttft === BigInt(-1) ? -1 : Math.trunc(Number(ttft / BigInt(1000000))),
+      durationMs,
       cancelled: this.abortController.signal.aborted,
       label: this.#llm.label(),
       completionTokens: usage?.completionTokens || 0,
       promptTokens: usage?.promptTokens || 0,
       promptCachedTokens: usage?.promptCachedTokens || 0,
       totalTokens: usage?.totalTokens || 0,
       tokensPerSecond: (() => {
-        const durationSeconds = Math.trunc(Number(duration / BigInt(1000000000)));
-        return durationSeconds > 0 ? (usage?.completionTokens || 0) / durationSeconds : 0;
+        if (durationMs <= 0) {
+          return 0;
+        }
+        return (usage?.completionTokens || 0) / (durationMs / 1000);
       })(),
     };
     this.#llm.emit('metrics_collected', metrics);
diff --git a/agents/src/metrics/base.ts b/agents/src/metrics/base.ts
@@ -15,8 +15,10 @@ export type LLMMetrics = {
   label: string;
   requestId: string;
   timestamp: number;
-  duration: number; // in milliseconds
-  ttft: number;
+  /** Duration of the request in milliseconds. */
+  durationMs: number;
+  /** Time to first token in milliseconds. */
+  ttftMs: number;
   cancelled: boolean;
   completionTokens: number;
   promptTokens: number;
@@ -32,13 +34,13 @@ export type STTMetrics = {
   requestId: string;
   timestamp: number;
   /**
-   * The request duration in seconds, 0.0 if the STT is streaming.
+   * The request duration in milliseconds, 0.0 if the STT is streaming.
    */
-  duration: number;
+  durationMs: number;
   /**
-   * The duration of the pushed audio in seconds.
+   * The duration of the pushed audio in milliseconds.
    */
-  audioDuration: number;
+  audioDurationMs: number;
   /**
    * Whether the STT is streaming (e.g using websocket).
    */
@@ -50,9 +52,12 @@ export type TTSMetrics = {
   label: string;
   requestId: string;
   timestamp: number;
-  ttfb: number;
-  duration: number;
-  audioDuration: number;
+  /** Time to first byte in milliseconds. */
+  ttfbMs: number;
+  /** Total synthesis duration in milliseconds. */
+  durationMs: number;
+  /** Generated audio duration in milliseconds. */
+  audioDurationMs: number;
   cancelled: boolean;
   charactersCount: number;
   streamed: boolean;
@@ -64,8 +69,8 @@ export type VADMetrics = {
   type: 'vad_metrics';
   label: string;
   timestamp: number;
-  idleTime: number;
-  inferenceDurationTotal: number;
+  idleTimeMs: number;
+  inferenceDurationTotalMs: number;
   inferenceCount: number;
 };
 
@@ -76,16 +81,16 @@ export type EOUMetrics = {
    * Amount of time between the end of speech from VAD and the decision to end the user's turn.
    * Set to 0.0 if the end of speech was not detected.
    */
-  endOfUtteranceDelay: number;
+  endOfUtteranceDelayMs: number;
   /**
    * Time taken to obtain the transcript after the end of the user's speech.
    * Set to 0.0 if the end of speech was not detected.
    */
-  transcriptionDelay: number;
+  transcriptionDelayMs: number;
   /**
    * Time taken to invoke the user's `Agent.onUserTurnCompleted` callback.
    */
-  onUserTurnCompletedDelay: number;
+  onUserTurnCompletedDelayMs: number;
   speechId?: string;
 };
 
@@ -118,13 +123,13 @@ export type RealtimeModelMetrics = {
    */
   timestamp: number;
   /**
-   * The duration of the response from created to done in seconds.
+   * The duration of the response from created to done in milliseconds.
    */
-  duration: number;
+  durationMs: number;
   /**
-   * Time to first audio token in seconds. -1 if no audio token was sent.
+   * Time to first audio token in milliseconds. -1 if no audio token was sent.
    */
-  ttft: number;
+  ttftMs: number;
   /**
    * Whether the request was cancelled.
    */
diff --git a/agents/src/metrics/usage_collector.ts b/agents/src/metrics/usage_collector.ts
@@ -8,7 +8,7 @@ export interface UsageSummary {
   llmPromptCachedTokens: number;
   llmCompletionTokens: number;
   ttsCharactersCount: number;
-  sttAudioDuration: number;
+  sttAudioDurationMs: number;
 }
 
 export class UsageCollector {
@@ -20,7 +20,7 @@ export class UsageCollector {
       llmPromptCachedTokens: 0,
       llmCompletionTokens: 0,
       ttsCharactersCount: 0,
-      sttAudioDuration: 0,
+      sttAudioDurationMs: 0,
     };
   }
 
@@ -36,7 +36,7 @@ export class UsageCollector {
     } else if (metrics.type === 'tts_metrics') {
       this.summary.ttsCharactersCount += metrics.charactersCount;
     } else if (metrics.type === 'stt_metrics') {
-      this.summary.sttAudioDuration += metrics.audioDuration;
+      this.summary.sttAudioDurationMs += metrics.audioDurationMs;
     }
   }
 
diff --git a/agents/src/metrics/utils.ts b/agents/src/metrics/utils.ts
@@ -13,7 +13,7 @@ export const logMetrics = (metrics: AgentMetrics) => {
   if (metrics.type === 'llm_metrics') {
     logger
       .child({
-        ttft: roundTwoDecimals(metrics.ttft),
+        ttftMs: roundTwoDecimals(metrics.ttftMs),
         inputTokens: metrics.promptTokens,
         promptCachedTokens: metrics.promptCachedTokens,
         outputTokens: metrics.completionTokens,
@@ -23,7 +23,7 @@ export const logMetrics = (metrics: AgentMetrics) => {
   } else if (metrics.type === 'realtime_model_metrics') {
     logger
       .child({
-        ttft: roundTwoDecimals(metrics.ttft),
+        ttftMs: roundTwoDecimals(metrics.ttftMs),
         input_tokens: metrics.inputTokens,
         cached_input_tokens: metrics.inputTokenDetails.cachedTokens,
         output_tokens: metrics.outputTokens,
@@ -34,21 +34,30 @@ export const logMetrics = (metrics: AgentMetrics) => {
   } else if (metrics.type === 'tts_metrics') {
     logger
       .child({
-        ttfb: roundTwoDecimals(metrics.ttfb),
-        audioDuration: metrics.audioDuration,
+        ttfbMs: roundTwoDecimals(metrics.ttfbMs),
+        audioDurationMs: Math.round(metrics.audioDurationMs),
       })
       .info('TTS metrics');
   } else if (metrics.type === 'eou_metrics') {
     logger
       .child({
-        end_of_utterance_delay: roundTwoDecimals(metrics.endOfUtteranceDelay),
-        transcription_delay: roundTwoDecimals(metrics.transcriptionDelay),
+        endOfUtteranceDelayMs: roundTwoDecimals(metrics.endOfUtteranceDelayMs),
+        transcriptionDelayMs: roundTwoDecimals(metrics.transcriptionDelayMs),
+        onUserTurnCompletedDelayMs: roundTwoDecimals(metrics.onUserTurnCompletedDelayMs),
       })
       .info('EOU metrics');
+  } else if (metrics.type === 'vad_metrics') {
+    logger
+      .child({
+        idleTimeMs: Math.round(metrics.idleTimeMs),
+        inferenceDurationTotalMs: Math.round(metrics.inferenceDurationTotalMs),
+        inferenceCount: metrics.inferenceCount,
+      })
+      .info('VAD metrics');
   } else if (metrics.type === 'stt_metrics') {
     logger
       .child({
-        audioDuration: metrics.audioDuration,
+        audioDurationMs: Math.round(metrics.audioDurationMs),
       })
       .info('STT metrics');
   }
diff --git a/agents/src/stt/stt.ts b/agents/src/stt/stt.ts
@@ -6,7 +6,7 @@ import type { TypedEventEmitter as TypedEmitter } from '@livekit/typed-emitter';
 import { EventEmitter } from 'node:events';
 import type { ReadableStream } from 'node:stream/web';
 import { APIConnectionError, APIError } from '../_exceptions.js';
-import { calculateAudioDuration } from '../audio.js';
+import { calculateAudioDurationSeconds } from '../audio.js';
 import { log } from '../log.js';
 import type { STTMetrics } from '../metrics/base.js';
 import { DeferredReadableStream } from '../stream/deferred_stream.js';
@@ -110,14 +110,14 @@ export abstract class STT extends (EventEmitter as new () => TypedEmitter<STTCal
   async recognize(frame: AudioBuffer): Promise<SpeechEvent> {
     const startTime = process.hrtime.bigint();
     const event = await this._recognize(frame);
-    const duration = Number((process.hrtime.bigint() - startTime) / BigInt(1000000));
+    const durationMs = Number((process.hrtime.bigint() - startTime) / BigInt(1000000));
     this.emit('metrics_collected', {
       type: 'stt_metrics',
       requestId: event.requestId ?? '',
       timestamp: Date.now(),
-      duration,
+      durationMs,
       label: this.label,
-      audioDuration: calculateAudioDuration(frame),
+      audioDurationMs: Math.round(calculateAudioDurationSeconds(frame) * 1000),
       streamed: false,
     });
     return event;
@@ -252,9 +252,9 @@ export abstract class SpeechStream implements AsyncIterableIterator<SpeechEvent>
         type: 'stt_metrics',
         timestamp: Date.now(),
         requestId: event.requestId!,
-        duration: 0,
+        durationMs: 0,
         label: this.#stt.label,
-        audioDuration: event.recognitionUsage!.audioDuration,
+        audioDurationMs: Math.round(event.recognitionUsage!.audioDuration * 1000),
         streamed: true,
       };
       this.#stt.emit('metrics_collected', metrics);
diff --git a/agents/src/tts/tts.ts b/agents/src/tts/tts.ts
@@ -228,22 +228,23 @@ export abstract class SynthesizeStream
 
   protected async monitorMetrics() {
     const startTime = process.hrtime.bigint();
-    let audioDuration = 0;
+    let audioDurationMs = 0;
     let ttfb: bigint = BigInt(-1);
     let requestId = '';
 
     const emit = () => {
       if (this.#metricsPendingTexts.length) {
         const text = this.#metricsPendingTexts.shift()!;
         const duration = process.hrtime.bigint() - startTime;
+        const roundedAudioDurationMs = Math.round(audioDurationMs);
         const metrics: TTSMetrics = {
           type: 'tts_metrics',
           timestamp: Date.now(),
           requestId,
-          ttfb: ttfb === BigInt(-1) ? -1 : Math.trunc(Number(ttfb / BigInt(1000000))),
-          duration: Math.trunc(Number(duration / BigInt(1000000))),
+          ttfbMs: ttfb === BigInt(-1) ? -1 : Math.trunc(Number(ttfb / BigInt(1000000))),
+          durationMs: Math.trunc(Number(duration / BigInt(1000000))),
           charactersCount: text.length,
-          audioDuration,
+          audioDurationMs: roundedAudioDurationMs,
           cancelled: this.abortController.signal.aborted,
           label: this.#tts.label,
           streamed: false,
@@ -263,7 +264,7 @@ export abstract class SynthesizeStream
         ttfb = process.hrtime.bigint() - startTime;
       }
       // TODO(AJS-102): use frame.durationMs once available in rtc-node
-      audioDuration += audio.frame.samplesPerChannel / audio.frame.sampleRate;
+      audioDurationMs += (audio.frame.samplesPerChannel / audio.frame.sampleRate) * 1000;
       if (audio.final) {
         emit();
       }
@@ -436,7 +437,7 @@ export abstract class ChunkedStream implements AsyncIterableIterator<Synthesized
 
   protected async monitorMetrics() {
     const startTime = process.hrtime.bigint();
-    let audioDuration = 0;
+    let audioDurationMs = 0;
     let ttfb: bigint = BigInt(-1);
     let requestId = '';
 
@@ -446,7 +447,7 @@ export abstract class ChunkedStream implements AsyncIterableIterator<Synthesized
       if (ttfb === BigInt(-1)) {
         ttfb = process.hrtime.bigint() - startTime;
       }
-      audioDuration += audio.frame.samplesPerChannel / audio.frame.sampleRate;
+      audioDurationMs += (audio.frame.samplesPerChannel / audio.frame.sampleRate) * 1000;
     }
     this.output.close();
 
@@ -455,10 +456,10 @@ export abstract class ChunkedStream implements AsyncIterableIterator<Synthesized
       type: 'tts_metrics',
       timestamp: Date.now(),
       requestId,
-      ttfb: ttfb === BigInt(-1) ? -1 : Math.trunc(Number(ttfb / BigInt(1000000))),
-      duration: Math.trunc(Number(duration / BigInt(1000000))),
+      ttfbMs: ttfb === BigInt(-1) ? -1 : Math.trunc(Number(ttfb / BigInt(1000000))),
+      durationMs: Math.trunc(Number(duration / BigInt(1000000))),
       charactersCount: this.#text.length,
-      audioDuration,
+      audioDurationMs: Math.round(audioDurationMs),
       cancelled: false, // TODO(AJS-186): support ChunkedStream with 1.0 - add this.abortController.signal.aborted here
       label: this.#tts.label,
       streamed: false,
diff --git a/agents/src/vad.ts b/agents/src/vad.ts
@@ -139,7 +139,7 @@ export abstract class VADStream implements AsyncIterableIterator<VADEvent> {
   }
 
   protected async monitorMetrics() {
-    let inferenceDurationTotal = 0;
+    let inferenceDurationTotalMs = 0;
     let inferenceCount = 0;
     const metricsReader = this.metricsStream.getReader();
     while (true) {
@@ -154,20 +154,20 @@ export abstract class VADStream implements AsyncIterableIterator<VADEvent> {
             this.#vad.emit('metrics_collected', {
               type: 'vad_metrics',
               timestamp: Date.now(),
-              idleTime: Math.trunc(
+              idleTimeMs: Math.trunc(
                 Number((process.hrtime.bigint() - this.#lastActivityTime) / BigInt(1000000)),
               ),
-              inferenceDurationTotal,
+              inferenceDurationTotalMs,
               inferenceCount,
               label: this.#vad.label,
             });
 
             inferenceCount = 0;
-            inferenceDurationTotal = 0;
+            inferenceDurationTotalMs = 0;
           }
           break;
         case VADEventType.INFERENCE_DONE:
-          inferenceDurationTotal += value.inferenceDuration;
+          inferenceDurationTotalMs += Math.round(value.inferenceDuration);
           this.#lastActivityTime = process.hrtime.bigint();
           break;
         case VADEventType.END_OF_SPEECH:
diff --git a/agents/src/voice/agent_activity.ts b/agents/src/voice/agent_activity.ts
@@ -984,9 +984,9 @@ export class AgentActivity implements RecognitionHooks {
     const eouMetrics: EOUMetrics = {
       type: 'eou_metrics',
       timestamp: Date.now(),
-      endOfUtteranceDelay: info.endOfUtteranceDelay,
-      transcriptionDelay: info.transcriptionDelay,
-      onUserTurnCompletedDelay: callbackDuration,
+      endOfUtteranceDelayMs: info.endOfUtteranceDelay,
+      transcriptionDelayMs: info.transcriptionDelay,
+      onUserTurnCompletedDelayMs: callbackDuration,
       speechId: speechHandle.id,
     };
 
diff --git a/plugins/google/src/beta/realtime/realtime_api.ts b/plugins/google/src/beta/realtime/realtime_api.ts
diff --git a/plugins/openai/src/realtime/realtime_model.ts b/plugins/openai/src/realtime/realtime_model.ts

Original file line number	Diff line number	Diff line change
`@@ -8,7 +8,7 @@ export interface UsageSummary {`
`8`	`8`	`llmPromptCachedTokens: number;`
`9`	`9`	`llmCompletionTokens: number;`
`10`	`10`	`ttsCharactersCount: number;`
`11`		`- sttAudioDuration: number;`
	`11`	`+ sttAudioDurationMs: number;`
`12`	`12`	`}`
`13`	`13`
`14`	`14`	`export class UsageCollector {`
`@@ -20,7 +20,7 @@ export class UsageCollector {`
`20`	`20`	`llmPromptCachedTokens: 0,`
`21`	`21`	`llmCompletionTokens: 0,`
`22`	`22`	`ttsCharactersCount: 0,`
`23`		`- sttAudioDuration: 0,`
	`23`	`+ sttAudioDurationMs: 0,`
`24`	`24`	`};`
`25`	`25`	`}`
`26`	`26`
`@@ -36,7 +36,7 @@ export class UsageCollector {`
`36`	`36`	`} else if (metrics.type === 'tts_metrics') {`
`37`	`37`	`this.summary.ttsCharactersCount += metrics.charactersCount;`
`38`	`38`	`} else if (metrics.type === 'stt_metrics') {`
`39`		`- this.summary.sttAudioDuration += metrics.audioDuration;`
	`39`	`+ this.summary.sttAudioDurationMs += metrics.audioDurationMs;`
`40`	`40`	`}`
`41`	`41`	`}`
`42`	`42`