@datalayer/agent-runtimes 1.5.42 → 1.5.44

package/lib/voice/serverHearing.js

@@ -6,7 +6,8 @@
  * Hearing on Datalayer's servers (VOICE.md VO-34, VO-14, §5): what was said
  * in one press, kept to its speech by Silero VAD in the page, then sent as a
  * WAV recording to the speech service's `POST /speech/transcriptions`, where
- * Whisper large-v3-turbo hears it — in any language it knows, and finding
+ * Whisper hears it — small for a spoken turn, large-v3-turbo past one
+ * (decision 17) — in any language it knows, and finding
  * the language itself when the person asks for that (detection, VO-14). The
  * sound is held in the request only: the service keeps neither it nor the
  * words (VO-60).
@@ -14,8 +15,15 @@
  * Used where the application's voice says `where: server`, where it says
  * `auto` and no model in the browser hears the language, and where the
  * person asked for their language to be detected — never when they chose to
- * keep speech on the device. A clip model: its words come at the end, not
- * while the person speaks (VO-11).
+ * keep speech on the device.
+ *
+ * While the person speaks, the words come too (VO-11): the first look at
+ * what was heard so far opens `WS /speech/transcriptions/stream` (the token
+ * in its first message, VO-37), the sound is sent as it is captured, 16-bit
+ * PCM at 16 kHz, and the service's partial words are what was heard so far,
+ * greyed in the composer until final. At the end of the press the stream is
+ * told so and its final words are the transcript. A press heard without a
+ * look before it (a short one) is sent whole as a recording instead.
  *
  * @module voice/serverHearing
  */
@@ -23,6 +31,12 @@ import { transcriberFor } from '../specs/voices';
 import { HEARING_RATE } from './capture';
 /** Where the speech service hears a recording, on ai-agents' host. */
 export const TRANSCRIPTIONS_PATH = '/api/ai-agents/v1/speech/transcriptions';
+/** Where the speech service hears speech as it is said. */
+export const TRANSCRIPTION_STREAM_PATH = '/api/ai-agents/v1/speech/transcriptions/stream';
+/** A stream no look or hearing has fed for this long is closed: its press was cancelled. */
+export const STREAM_IDLE_MS = 3000;
+/** How long the final words of a stream are waited for. */
+export const STREAM_FINAL_MS = 60_000;
 /** The model the speech service hears with (agentspecs' catalogue). */
 export const SERVER_HEARING_MODEL = 'whisper-large-v3-turbo';
 /** The shortest speech worth hearing, in samples: a word said alone. */
@@ -72,6 +86,162 @@ export function wavOf(samples, rate = HEARING_RATE) {
     }
     return bytes;
 }
+/** 16 kHz samples in [-1, 1] as little-endian 16-bit PCM: what a `pcm16` stream takes. */
+export function pcm16Of(samples) {
+    const bytes = new ArrayBuffer(samples.length * 2);
+    const view = new DataView(bytes);
+    for (let i = 0; i < samples.length; i++) {
+        const sample = Math.max(-1, Math.min(1, samples[i]));
+        view.setInt16(i * 2, sample < 0 ? sample * 0x8000 : sample * 0x7fff, true);
+    }
+    return bytes;
+}
+/** The ws(s) address of the stream on ai-agents' origin. */
+export function streamUrlOf(speechUrl) {
+    return `${speechUrl.replace(/\/+$/, '').replace(/^http/, 'ws')}${TRANSCRIPTION_STREAM_PATH}`;
+}
+/**
+ * One press heard as it is said: the sound sent as it comes, the service's
+ * partial words kept, its final words awaited at the end.
+ */
+class PressStream {
+    /** Samples of the press sent so far. */
+    sent = 0;
+    /** The newest partial words. */
+    partial = '';
+    /** Why the service refused, in its sentence. */
+    refused;
+    socket;
+    ready = false;
+    queued = [];
+    final;
+    waiting;
+    closed = false;
+    idle;
+    constructor(open, speechUrl, first) {
+        this.socket = open(streamUrlOf(speechUrl));
+        this.socket.binaryType = 'arraybuffer';
+        this.socket.onopen = () => {
+            this.socket.send(JSON.stringify({
+                token: first.token ?? '',
+                language: first.language ?? '',
+                format: 'pcm16',
+            }));
+        };
+        this.socket.onmessage = event => {
+            let said;
+            try {
+                said = JSON.parse(String(event.data));
+            }
+            catch {
+                return;
+            }
+            if (said.type === 'ready') {
+                this.ready = true;
+                for (const chunk of this.queued.splice(0)) {
+                    this.socket.send(chunk);
+                }
+            }
+            else if (said.type === 'partial' && typeof said.text === 'string') {
+                this.partial = said.text.trim();
+            }
+            else if (said.type === 'final') {
+                this.final = {
+                    text: typeof said.text === 'string' ? said.text.trim() : '',
+                    language: typeof said.language === 'string' ? said.language : '',
+                };
+                this.settle();
+            }
+            else if (said.type === 'refused') {
+                this.refused =
+                    typeof said.detail === 'string' && said.detail
+                        ? said.detail
+                        : "What you said could not be heard on Datalayer's servers.";
+                this.settle();
+            }
+        };
+        this.socket.onclose = () => {
+            this.closed = true;
+            this.settle();
+        };
+        this.socket.onerror = () => undefined;
+        this.touch();
+    }
+    /** Send what the press added since the last call. */
+    feed(audio) {
+        this.touch();
+        if (audio.length <= this.sent || this.closed) {
+            return;
+        }
+        const chunk = pcm16Of(audio.subarray(this.sent));
+        this.sent = audio.length;
+        if (this.ready) {
+            this.socket.send(chunk);
+        }
+        else {
+            this.queued.push(chunk);
+        }
+    }
+    /** The end of the press: the final words, or `undefined` when the stream ended without them. */
+    finish(audio) {
+        this.feed(audio);
+        if (this.final || this.closed || this.refused) {
+            return Promise.resolve(this.final);
+        }
+        return new Promise(resolve => {
+            const timer = setTimeout(() => {
+                this.waiting = undefined;
+                this.close();
+                resolve(undefined);
+            }, STREAM_FINAL_MS);
+            this.waiting = final => {
+                clearTimeout(timer);
+                resolve(final);
+            };
+            const end = JSON.stringify({ type: 'end' });
+            if (this.ready) {
+                this.socket.send(end);
+            }
+            else {
+                // Said once the service is ready, after the sound queued before it.
+                this.queued.push(end);
+            }
+        });
+    }
+    close() {
+        if (this.idle) {
+            clearTimeout(this.idle);
+        }
+        this.closed = true;
+        try {
+            this.socket.close(1000);
+        }
+        catch {
+            // Already closed.
+        }
+        this.settle();
+    }
+    touch() {
+        if (this.idle) {
+            clearTimeout(this.idle);
+        }
+        this.idle = setTimeout(() => {
+            if (!this.waiting) {
+                this.close();
+            }
+        }, STREAM_IDLE_MS);
+    }
+    settle() {
+        const waiting = this.waiting;
+        this.waiting = undefined;
+        waiting?.(this.final);
+        if (this.final || this.refused) {
+            if (this.idle) {
+                clearTimeout(this.idle);
+            }
+        }
+    }
+}
 /**
  * Hears on Datalayer's servers: the browser's VAD finds the speech and when
  * it stopped; Whisper, on the speech service, writes it.
@@ -88,17 +258,75 @@ export class ServerHearing {
     ready(_language, progress) {
         return this.finder.readyToFindSpeech(progress);
     }
-    /** Whether speech was found and how long it has been silent; its words come at the end. */
-    async hearSoFar(audio) {
+    /** The press being heard as it is said, while one is. */
+    press;
+    /** The press's stream, opened anew when a new press began (its sound is shorter). */
+    streamFor(audio, language) {
+        const open = this.options.socket ??
+            (typeof WebSocket === 'undefined'
+                ? undefined
+                : (url) => new WebSocket(url));
+        if (!open) {
+            return undefined;
+        }
+        if (this.press && audio.length < this.press.sent) {
+            this.press.close();
+            this.press = undefined;
+        }
+        if (!this.press) {
+            this.press = new PressStream(open, this.options.speechUrl, {
+                token: this.options.token,
+                language: this.options.detect ? '' : language,
+            });
+        }
+        return this.press;
+    }
+    /**
+     * Whether speech was found and how long it has been silent, and the words
+     * the speech service heard so far: the sound goes to its stream as it comes.
+     */
+    async hearSoFar(audio, language = '') {
         const { speech, silentMs } = await this.finder.findSpeech(audio);
-        return { text: '', spoke: speech.length >= SHORTEST, silentMs };
+        const spoke = speech.length >= SHORTEST;
+        const press = this.streamFor(audio, language);
+        press?.feed(audio);
+        return { text: spoke ? (press?.partial ?? '') : '', spoke, silentMs };
     }
     async hear(audio, language) {
         const started = performance.now();
+        const press = this.press && audio.length >= this.press.sent ? this.press : undefined;
+        if (this.press && !press) {
+            this.press.close();
+        }
+        this.press = undefined;
         const { speech } = await this.finder.findSpeech(audio);
         if (speech.length < SHORTEST) {
+            press?.close();
             return undefined;
         }
+        if (press) {
+            const final = await press.finish(audio);
+            press.close();
+            if (press.refused) {
+                throw new Error(press.refused);
+            }
+            if (!final) {
+                throw new Error("What you said could not be heard on Datalayer's servers: the stream ended before its words.");
+            }
+            return final.text
+                ? {
+                    text: final.text,
+                    metadata: {
+                        input: 'voice',
+                        language: final.language || language,
+                        engine: SERVER_HEARING_MODEL,
+                        where: 'server',
+                    },
+                    seconds: audio.length / HEARING_RATE,
+                    ms: Math.round(performance.now() - started),
+                }
+                : undefined;
+        }
         const { speechUrl, token, detect, fetcher = fetch } = this.options;
         const query = detect ? '' : `?language=${encodeURIComponent(language)}`;
         const answered = await fetcher(`${speechUrl.replace(/\/+$/, '')}${TRANSCRIPTIONS_PATH}${query}`, {

package/lib/voice/settings.d.ts

@@ -33,6 +33,16 @@ export declare const VOLUME_RANGE: {
     readonly max: 1;
     readonly default: 1;
 };
+/**
+ * How much longer than usual the person pauses before they are done
+ * (VOICE.md VO-69): the silence that stops a microphone clicked on, and the
+ * one that ends a hands-free turn, times this.
+ */
+export declare const PAUSE_RANGE: {
+    readonly min: 1;
+    readonly max: 4;
+    readonly default: 1;
+};
 /**
  * What a person chose for themselves. Every field is optional: an absent one
  * means "whatever the application says".
@@ -62,6 +72,8 @@ export interface VoiceSettings {
      * heard on Datalayer's servers, which detect it, rather than in the browser.
      */
     detectLanguage?: boolean;
+    /** How much longer they pause before they are done (VO-69): 1 is usual. */
+    pause?: number;
 }
 /** The settings as they can be used: numbers in range, nothing else trusted. */
 export declare function saneVoiceSettings(settings: VoiceSettings | null | undefined): VoiceSettings;

package/lib/voice/settings.js

@@ -7,6 +7,12 @@ export const VOICE_SETTINGS_KEY = 'datalayer-voice-settings';
 /** How fast and how loud, as multipliers of the voice's own. */
 export const RATE_RANGE = { min: 0.5, max: 2, default: 1 };
 export const VOLUME_RANGE = { min: 0, max: 1, default: 1 };
+/**
+ * How much longer than usual the person pauses before they are done
+ * (VOICE.md VO-69): the silence that stops a microphone clicked on, and the
+ * one that ends a hands-free turn, times this.
+ */
+export const PAUSE_RANGE = { min: 1, max: 4, default: 1 };
 function clamp(value, { min, max }) {
     return Math.min(max, Math.max(min, value));
 }
@@ -43,6 +49,9 @@ export function saneVoiceSettings(settings) {
     if (typeof kept.detectLanguage === 'boolean') {
         sane.detectLanguage = kept.detectLanguage;
     }
+    if (typeof kept.pause === 'number' && Number.isFinite(kept.pause)) {
+        sane.pause = clamp(kept.pause, PAUSE_RANGE);
+    }
     return sane;
 }
 /** What was kept, or nothing. Storage may be refused, or hold nonsense. */

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@datalayer/agent-runtimes",
-    "version": "1.5.42",
+    "version": "1.5.44",
     "type": "module",
     "workspaces": [
         ".",