@burdenoff/website-sdk 2026.1105.2 → 2026.1105.3

package/dist/avatar-mic/index.d.mts

@@ -1,62 +1,3 @@
-/** What the worklet posts for each frame. */
-interface CaptureFrameMessage {
-    /** s16le mono 16 kHz, `CAPTURE_FRAME_SAMPLES` samples. */
-    pcm: ArrayBuffer;
-    /** Microphone level over the frame, dBFS. */
-    rmsDb: number;
-    /** Reference-bus level over the same samples, dBFS (about -90 when silent, as it always is here). */
-    refDb: number;
-    /** Zero crossings per sample. */
-    zcr: number;
-    /** AudioContext time at the end of the frame, seconds. */
-    ctxTime: number;
-    seq: number;
-}
-
-/**
- * BOFF-9797 (Stage C). The Avatar tab's own microphone capture: `getUserMedia`, through the capture worklet
- * (`capture-worklet.ts`), into 16 kHz PCM frames.
- *
- * **Ported, not shared.** From `src/shared/assistant/voice/handsfree/micCapture.ts` in Algoshred/fe-libs at the
- * commit `capture-worklet.ts` names (`CAPTURE_WORKLET_PORTED_FROM`; blob `8486c630`), with one difference: fe-libs
- * routes everything its assistant plays into the worklet's second input (its playback reference, for barge-in), and the
- * website plays nothing through this context (the avatar's voice is the room's own audio element, and the mic is
- * offered only while no session plays), so no reference bus is made and the worklet has the microphone as its only
- * input. `looksLikeHeadset`, which only fe-libs' barge-in reads, is not ported. `capture-worklet.test.ts` holds the
- * parity tests.
- *
- * It opens `getUserMedia` with echo cancellation, noise suppression and auto gain on. Constraints are plain `true`,
- * never `{ exact }`: `exact` throws OverconstrainedError on handsets that cannot honour one. The settings the browser
- * actually applied are read back (`track.getSettings()`), because some Android handsets report `true` and apply nothing.
- *
- * `stop()` stops the track (the browser's recording indicator goes), disconnects the graph and closes the context. The
- * caller does that the moment a press ends, whichever way it ends; `pagehide` does it too, whatever the caller does.
- */
-
-interface MicCaptureHandle {
-    context: AudioContext;
-    micSource: MediaStreamAudioSourceNode;
-    track: MediaStreamTrack;
-    /** What the browser reports it applied. */
-    settings: MediaTrackSettings;
-    /** Stop everything and release the microphone. Safe to call more than once. */
-    stop(): void;
-}
-
-/** What the mic needs of a WebSocket; {@link openBrowserSocket} adapts the browser's. (fe-libs' seam.) */
-interface SocketLike {
-    readonly readyState: number;
-    send(data: string | ArrayBufferLike | ArrayBufferView): void;
-    close(code?: number): void;
-    listen(handlers: SocketHandlers): void;
-}
-interface SocketHandlers {
-    open(): void;
-    message(data: unknown): void;
-    close(code: number): void;
-    error(): void;
-}
-
 /**
  * BOFF-9797 (Stage C). The shapes the Avatar tab's mic shares with the live avatar's view.
  *
@@ -79,7 +20,8 @@ interface AvatarMicOffer {
  * Why a press did not give text:
  *
  * - `denied`: the microphone was refused (the visitor, the browser or the page's permissions);
- * - `unsupported`: this browser cannot capture here, or has no microphone to use;
+ * - `unsupported`: this browser cannot capture here, or has no microphone to use (with the browser's recognizer: its
+ *   speech service is off or unreachable here, or it has no language to offer);
  * - `unavailable`: the server will not take a stream from this page now: no offer, an offer it refuses twice, its mic
  *   notice is another version than this SDK's, or the website avatar is paused;
  * - `limit`: a limit on voice questions is reached (the visitor's, the network's, or every visitor's together);
@@ -151,9 +93,191 @@ interface AvatarMic {
     /** The press as it stands. */
     state(): AvatarMicState;
 }
+/**
+ * The mic through the browser's own speech recognition (Stage C, `mic.mode: "browser"` in availability): the browser
+ * turns the voice into text, and nothing of the press reaches the server. Same press, same states, same failures as
+ * the stream's ({@link AvatarMic}, {@link AvatarMicState}); it never reports `unavailable`, `limit` or `busy`.
+ */
+interface BrowserAvatarMicOptions {
+    /** The recognizer's language, a BCP-47 tag (`en-IN`). */
+    lang: string;
+    /** Tried once, in the same press, when the browser does not recognise `lang` (`en-US`). */
+    fallbackLang: string;
+    /** The longest press, from the moment the browser starts listening: thirty seconds. */
+    maxPressMs: number;
+    /** Every change of the press's state. */
+    onState: (state: AvatarMicState) => void;
+}
 /** What the mic's chunk (`dist/avatar-mic/`) exports. */
 interface AvatarMicModule {
     createAvatarMic: (options: AvatarMicOptions) => AvatarMic;
+    createBrowserAvatarMic: (options: BrowserAvatarMicOptions) => AvatarMic;
+}
+
+/**
+ * BOFF-9797 (Stage C, the browser's own recognizer). The Avatar tab's mic through the browser's speech recognition
+ * (`SpeechRecognition`, `webkitSpeechRecognition`), as the product assistant listens: tap to start, tap to stop,
+ * thirty seconds at most. The browser turns the voice into text (its own speech service: Google's for Chrome, Apple's
+ * for Safari, Microsoft's for Edge); this module never sends anything anywhere, and the text goes only to the view,
+ * which puts it in the box for the visitor to check and Send.
+ *
+ * The phones' quirks are the ones the SDK's Chat dictation (`src/hooks/use-speech-input.ts`, live on the sites) met
+ * there, handled the same way: the permission prompt raised by opening the microphone first, a pause before the
+ * recognizer opens its own, one quiet retry after a fresh grant, and one utterance per recognizer on a touch device.
+ *
+ * A press, in order:
+ *
+ * 1. **The tap** (`start`, called inside the tap where the view can). Where the browser reports the microphone granted
+ *    (or reports nothing, and a press on this page has had it), the recognizer starts inside the tap. Otherwise the
+ *    microphone is opened and released at once, inside the tap (`getUserMedia`): that is what raises the browser's
+ *    permission prompt on phones, where a recognizer's own `start()` is refused (`not-allowed`) without asking anyone.
+ *    Nothing of that capture is kept, played or sent. Once it is granted, the recognizer starts
+ *    {@link BROWSER_PRIME_SETTLE_MS} later, so it does not race the released capture for the microphone. The
+ *    recognizer: interim results, one alternative, the press's language, and `continuous`, except on a touch device,
+ *    whose recognizer hears one utterance whatever it is asked (and, asked for more, fails sooner). A press still
+ *    finishing is superseded (aborted, its text never reported; the view lets no tap through meanwhile).
+ * 2. **The browser starts listening** (`onstart`, the honest signal: never while a prompt is up): `listening`, and the
+ *    press's limit counts from here ({@link BrowserAvatarMicOptions.maxPressMs}).
+ * 3. **Words as they are recognised**: rebuilt from the whole result list each time (finals, then the words still
+ *    being recognised), so a result sent again never doubles up, and a final that repeats the one before it with more
+ *    words replaces it (Android's cumulative finals).
+ * 4. **The engine ending by itself** (a pause, a phone's one utterance, a quiet spell's `no-speech`) does NOT end the
+ *    press: tap to stop holds on every platform. A new recognizer takes over in the same press, its words after the
+ *    ones before: at once on a computer, after {@link BROWSER_TOUCH_RESTART_MS} on a touch device, and after
+ *    {@link BROWSER_RESTART_COOLDOWN_MS} when one has just thrown on `start()` or, on a computer, ended as it began.
+ *    Bounded against a recognizer that never listens: after {@link BROWSER_MAX_QUICK_ENDS} ends in a row as they
+ *    began, or {@link BROWSER_MAX_RESTARTS} recognizers in all, the press gives what it has. Events of a recognizer that
+ *    has been replaced are dropped.
+ * 5. **The stop** (the visitor's tap, or the limit): `stop()`, not `abort()`, as browsers deliver the last words after
+ *    it; up to {@link BROWSER_FINAL_WAIT_MS} for its end. A tap while it is still starting (the prompt included) calls
+ *    the press off.
+ * 6. **The text** (`done`): what was heard, final or not (an engine that never marks a result final, iOS Safari, still
+ *    gives its words: the visitor checks them in the box anyway). Empty when nothing was heard.
+ *
+ * **Released on every way out.** `cancel()` (the view going, the page hidden, the Chat tab), a failure, a supersede and
+ * the end of a press all detach the recognizer's handlers first and then `abort()` it, so the browser's recording
+ * indicator goes and nothing it says later is reported. A microphone opened for the prompt is released as soon as it
+ * opens, whether or not the press is still there.
+ *
+ * Errors. Opening the microphone for the prompt: refused (`NotAllowedError`, `SecurityError`) is `denied`; no
+ * microphone (`NotFoundError`) is `unsupported`; anything else is `failed`. The recognizer's
+ * (`SpeechRecognitionErrorEvent.error`): `not-allowed` or `service-not-allowed` while the microphone is confirmed for
+ * the press (the browser reports it granted, or opening it just worked) is tried once more after
+ * {@link BROWSER_GRANT_RETRY_MS}, quietly (Android refuses the first start after a fresh grant and takes the next), and
+ * after that it is `unsupported`: the speech service refusing, not a permission the visitor can give. Unconfirmed,
+ * `not-allowed` is `denied` (and the next tap asks the browser again) and `service-not-allowed` (iOS with Siri and
+ * Dictation off, Edge's policy) is `unsupported`; `audio-capture` (no microphone) is `unsupported`;
+ * `language-not-supported` tries the fallback language once, then `unsupported`; `network` is `failed`, and
+ * `unsupported` the second time running on the page; `no-speech` is a quiet spell (a new recognizer); anything else is
+ * `failed`. A failure once words were heard, whatever it is, gives those words instead (`done`).
+ */
+
+/** One alternative of a recognition result: what the browser heard. */
+interface RecognitionAlternativeLike {
+    readonly transcript: string;
+}
+/** One recognition result: final, or still being recognised. */
+interface RecognitionResultLike {
+    readonly isFinal: boolean;
+    readonly length: number;
+    readonly [index: number]: RecognitionAlternativeLike | undefined;
+}
+/** The whole result list of a recognizer, as each `result` event carries it. */
+interface RecognitionResultListLike {
+    readonly length: number;
+    readonly [index: number]: RecognitionResultLike | undefined;
+}
+/** What this module uses of a browser's `SpeechRecognition` (declared here: TypeScript's DOM library has none). */
+interface RecognitionLike {
+    lang: string;
+    continuous: boolean;
+    interimResults: boolean;
+    maxAlternatives: number;
+    onstart: (() => void) | null;
+    onresult: ((event: {
+        readonly results: RecognitionResultListLike;
+    }) => void) | null;
+    onerror: ((event: {
+        readonly error: string;
+    }) => void) | null;
+    onend: (() => void) | null;
+    start(): void;
+    stop(): void;
+    abort(): void;
+}
+/** What a press needs from the browser. The browser's own by default; a test hands in fakes. */
+interface BrowserAvatarMicSeams {
+    /** A new recognizer, or null where the browser has none. */
+    createRecognition: () => RecognitionLike | null;
+    /** A monotonic clock, in milliseconds, for the cooldowns. */
+    now: () => number;
+    /** {@link primeBrowserMicrophone}: called inside the tap, where the microphone is not known to be granted. */
+    primeMicrophone: () => Promise<void> | null;
+    /** {@link watchBrowserMicPermission}: called once, as the mic is made. */
+    watchMicPermission: (onState: (state: PermissionState) => void) => void;
+    /** {@link hasCoarsePointer}: read at each tap. */
+    touch: () => boolean;
+}
+/** The Avatar tab's mic through the browser's own recognizer. See the file's comment for what a press does. */
+declare function createBrowserAvatarMic(options: BrowserAvatarMicOptions, seams?: Partial<BrowserAvatarMicSeams>): AvatarMic;
+
+/** What the worklet posts for each frame. */
+interface CaptureFrameMessage {
+    /** s16le mono 16 kHz, `CAPTURE_FRAME_SAMPLES` samples. */
+    pcm: ArrayBuffer;
+    /** Microphone level over the frame, dBFS. */
+    rmsDb: number;
+    /** Reference-bus level over the same samples, dBFS (about -90 when silent, as it always is here). */
+    refDb: number;
+    /** Zero crossings per sample. */
+    zcr: number;
+    /** AudioContext time at the end of the frame, seconds. */
+    ctxTime: number;
+    seq: number;
+}
+
+/**
+ * BOFF-9797 (Stage C). The Avatar tab's own microphone capture: `getUserMedia`, through the capture worklet
+ * (`capture-worklet.ts`), into 16 kHz PCM frames.
+ *
+ * **Ported, not shared.** From `src/shared/assistant/voice/handsfree/micCapture.ts` in Algoshred/fe-libs at the
+ * commit `capture-worklet.ts` names (`CAPTURE_WORKLET_PORTED_FROM`; blob `8486c630`), with one difference: fe-libs
+ * routes everything its assistant plays into the worklet's second input (its playback reference, for barge-in), and the
+ * website plays nothing through this context (the avatar's voice is the room's own audio element, and the mic is
+ * offered only while no session plays), so no reference bus is made and the worklet has the microphone as its only
+ * input. `looksLikeHeadset`, which only fe-libs' barge-in reads, is not ported. `capture-worklet.test.ts` holds the
+ * parity tests.
+ *
+ * It opens `getUserMedia` with echo cancellation, noise suppression and auto gain on. Constraints are plain `true`,
+ * never `{ exact }`: `exact` throws OverconstrainedError on handsets that cannot honour one. The settings the browser
+ * actually applied are read back (`track.getSettings()`), because some Android handsets report `true` and apply nothing.
+ *
+ * `stop()` stops the track (the browser's recording indicator goes), disconnects the graph and closes the context. The
+ * caller does that the moment a press ends, whichever way it ends; `pagehide` does it too, whatever the caller does.
+ */
+
+interface MicCaptureHandle {
+    context: AudioContext;
+    micSource: MediaStreamAudioSourceNode;
+    track: MediaStreamTrack;
+    /** What the browser reports it applied. */
+    settings: MediaTrackSettings;
+    /** Stop everything and release the microphone. Safe to call more than once. */
+    stop(): void;
+}
+
+/** What the mic needs of a WebSocket; {@link openBrowserSocket} adapts the browser's. (fe-libs' seam.) */
+interface SocketLike {
+    readonly readyState: number;
+    send(data: string | ArrayBufferLike | ArrayBufferView): void;
+    close(code?: number): void;
+    listen(handlers: SocketHandlers): void;
+}
+interface SocketHandlers {
+    open(): void;
+    message(data: unknown): void;
+    close(code: number): void;
+    error(): void;
 }
 
 /**
@@ -176,7 +300,8 @@ interface AvatarMicModule {
  *
  * **Released on every way out.** The microphone's track is stopped the moment the recording stops, and also when the
  * press fails, is cancelled (`cancel()`: the view going, the page hidden, the Chat tab) or superseded (a new press while
- * one is still finishing: `cancel` for the old one, so it stops being paid for, and its text is never reported). A
+ * one is still finishing, a tap the view lets through for this stream mic only: `cancel` for the old one, so it stops
+ * being paid for, and its text is never reported). A
  * microphone that opens after its press has gone is closed as it opens. The stream is closed whenever no press needs
  * it: one connection serves a press and the press that supersedes it, and no other.
  */
@@ -190,4 +315,4 @@ interface AvatarMicSeams {
 /** The mic of one Avatar tab. See the file's comment for what a press does. */
 declare function createAvatarMic(options: AvatarMicOptions, seams?: Partial<AvatarMicSeams>): AvatarMic;
 
-export { type AvatarMic, type AvatarMicFailure, type AvatarMicModule, type AvatarMicOffer, type AvatarMicOptions, type AvatarMicState, createAvatarMic };
+export { type AvatarMic, type AvatarMicFailure, type AvatarMicModule, type AvatarMicOffer, type AvatarMicOptions, type AvatarMicState, type BrowserAvatarMicOptions, createAvatarMic, createBrowserAvatarMic };

package/dist/avatar-mic/index.d.ts

@@ -1,62 +1,3 @@
-/** What the worklet posts for each frame. */
-interface CaptureFrameMessage {
-    /** s16le mono 16 kHz, `CAPTURE_FRAME_SAMPLES` samples. */
-    pcm: ArrayBuffer;
-    /** Microphone level over the frame, dBFS. */
-    rmsDb: number;
-    /** Reference-bus level over the same samples, dBFS (about -90 when silent, as it always is here). */
-    refDb: number;
-    /** Zero crossings per sample. */
-    zcr: number;
-    /** AudioContext time at the end of the frame, seconds. */
-    ctxTime: number;
-    seq: number;
-}
-
-/**
- * BOFF-9797 (Stage C). The Avatar tab's own microphone capture: `getUserMedia`, through the capture worklet
- * (`capture-worklet.ts`), into 16 kHz PCM frames.
- *
- * **Ported, not shared.** From `src/shared/assistant/voice/handsfree/micCapture.ts` in Algoshred/fe-libs at the
- * commit `capture-worklet.ts` names (`CAPTURE_WORKLET_PORTED_FROM`; blob `8486c630`), with one difference: fe-libs
- * routes everything its assistant plays into the worklet's second input (its playback reference, for barge-in), and the
- * website plays nothing through this context (the avatar's voice is the room's own audio element, and the mic is
- * offered only while no session plays), so no reference bus is made and the worklet has the microphone as its only
- * input. `looksLikeHeadset`, which only fe-libs' barge-in reads, is not ported. `capture-worklet.test.ts` holds the
- * parity tests.
- *
- * It opens `getUserMedia` with echo cancellation, noise suppression and auto gain on. Constraints are plain `true`,
- * never `{ exact }`: `exact` throws OverconstrainedError on handsets that cannot honour one. The settings the browser
- * actually applied are read back (`track.getSettings()`), because some Android handsets report `true` and apply nothing.
- *
- * `stop()` stops the track (the browser's recording indicator goes), disconnects the graph and closes the context. The
- * caller does that the moment a press ends, whichever way it ends; `pagehide` does it too, whatever the caller does.
- */
-
-interface MicCaptureHandle {
-    context: AudioContext;
-    micSource: MediaStreamAudioSourceNode;
-    track: MediaStreamTrack;
-    /** What the browser reports it applied. */
-    settings: MediaTrackSettings;
-    /** Stop everything and release the microphone. Safe to call more than once. */
-    stop(): void;
-}
-
-/** What the mic needs of a WebSocket; {@link openBrowserSocket} adapts the browser's. (fe-libs' seam.) */
-interface SocketLike {
-    readonly readyState: number;
-    send(data: string | ArrayBufferLike | ArrayBufferView): void;
-    close(code?: number): void;
-    listen(handlers: SocketHandlers): void;
-}
-interface SocketHandlers {
-    open(): void;
-    message(data: unknown): void;
-    close(code: number): void;
-    error(): void;
-}
-
 /**
  * BOFF-9797 (Stage C). The shapes the Avatar tab's mic shares with the live avatar's view.
  *
@@ -79,7 +20,8 @@ interface AvatarMicOffer {
  * Why a press did not give text:
  *
  * - `denied`: the microphone was refused (the visitor, the browser or the page's permissions);
- * - `unsupported`: this browser cannot capture here, or has no microphone to use;
+ * - `unsupported`: this browser cannot capture here, or has no microphone to use (with the browser's recognizer: its
+ *   speech service is off or unreachable here, or it has no language to offer);
  * - `unavailable`: the server will not take a stream from this page now: no offer, an offer it refuses twice, its mic
  *   notice is another version than this SDK's, or the website avatar is paused;
  * - `limit`: a limit on voice questions is reached (the visitor's, the network's, or every visitor's together);
@@ -151,9 +93,191 @@ interface AvatarMic {
     /** The press as it stands. */
     state(): AvatarMicState;
 }
+/**
+ * The mic through the browser's own speech recognition (Stage C, `mic.mode: "browser"` in availability): the browser
+ * turns the voice into text, and nothing of the press reaches the server. Same press, same states, same failures as
+ * the stream's ({@link AvatarMic}, {@link AvatarMicState}); it never reports `unavailable`, `limit` or `busy`.
+ */
+interface BrowserAvatarMicOptions {
+    /** The recognizer's language, a BCP-47 tag (`en-IN`). */
+    lang: string;
+    /** Tried once, in the same press, when the browser does not recognise `lang` (`en-US`). */
+    fallbackLang: string;
+    /** The longest press, from the moment the browser starts listening: thirty seconds. */
+    maxPressMs: number;
+    /** Every change of the press's state. */
+    onState: (state: AvatarMicState) => void;
+}
 /** What the mic's chunk (`dist/avatar-mic/`) exports. */
 interface AvatarMicModule {
     createAvatarMic: (options: AvatarMicOptions) => AvatarMic;
+    createBrowserAvatarMic: (options: BrowserAvatarMicOptions) => AvatarMic;
+}
+
+/**
+ * BOFF-9797 (Stage C, the browser's own recognizer). The Avatar tab's mic through the browser's speech recognition
+ * (`SpeechRecognition`, `webkitSpeechRecognition`), as the product assistant listens: tap to start, tap to stop,
+ * thirty seconds at most. The browser turns the voice into text (its own speech service: Google's for Chrome, Apple's
+ * for Safari, Microsoft's for Edge); this module never sends anything anywhere, and the text goes only to the view,
+ * which puts it in the box for the visitor to check and Send.
+ *
+ * The phones' quirks are the ones the SDK's Chat dictation (`src/hooks/use-speech-input.ts`, live on the sites) met
+ * there, handled the same way: the permission prompt raised by opening the microphone first, a pause before the
+ * recognizer opens its own, one quiet retry after a fresh grant, and one utterance per recognizer on a touch device.
+ *
+ * A press, in order:
+ *
+ * 1. **The tap** (`start`, called inside the tap where the view can). Where the browser reports the microphone granted
+ *    (or reports nothing, and a press on this page has had it), the recognizer starts inside the tap. Otherwise the
+ *    microphone is opened and released at once, inside the tap (`getUserMedia`): that is what raises the browser's
+ *    permission prompt on phones, where a recognizer's own `start()` is refused (`not-allowed`) without asking anyone.
+ *    Nothing of that capture is kept, played or sent. Once it is granted, the recognizer starts
+ *    {@link BROWSER_PRIME_SETTLE_MS} later, so it does not race the released capture for the microphone. The
+ *    recognizer: interim results, one alternative, the press's language, and `continuous`, except on a touch device,
+ *    whose recognizer hears one utterance whatever it is asked (and, asked for more, fails sooner). A press still
+ *    finishing is superseded (aborted, its text never reported; the view lets no tap through meanwhile).
+ * 2. **The browser starts listening** (`onstart`, the honest signal: never while a prompt is up): `listening`, and the
+ *    press's limit counts from here ({@link BrowserAvatarMicOptions.maxPressMs}).
+ * 3. **Words as they are recognised**: rebuilt from the whole result list each time (finals, then the words still
+ *    being recognised), so a result sent again never doubles up, and a final that repeats the one before it with more
+ *    words replaces it (Android's cumulative finals).
+ * 4. **The engine ending by itself** (a pause, a phone's one utterance, a quiet spell's `no-speech`) does NOT end the
+ *    press: tap to stop holds on every platform. A new recognizer takes over in the same press, its words after the
+ *    ones before: at once on a computer, after {@link BROWSER_TOUCH_RESTART_MS} on a touch device, and after
+ *    {@link BROWSER_RESTART_COOLDOWN_MS} when one has just thrown on `start()` or, on a computer, ended as it began.
+ *    Bounded against a recognizer that never listens: after {@link BROWSER_MAX_QUICK_ENDS} ends in a row as they
+ *    began, or {@link BROWSER_MAX_RESTARTS} recognizers in all, the press gives what it has. Events of a recognizer that
+ *    has been replaced are dropped.
+ * 5. **The stop** (the visitor's tap, or the limit): `stop()`, not `abort()`, as browsers deliver the last words after
+ *    it; up to {@link BROWSER_FINAL_WAIT_MS} for its end. A tap while it is still starting (the prompt included) calls
+ *    the press off.
+ * 6. **The text** (`done`): what was heard, final or not (an engine that never marks a result final, iOS Safari, still
+ *    gives its words: the visitor checks them in the box anyway). Empty when nothing was heard.
+ *
+ * **Released on every way out.** `cancel()` (the view going, the page hidden, the Chat tab), a failure, a supersede and
+ * the end of a press all detach the recognizer's handlers first and then `abort()` it, so the browser's recording
+ * indicator goes and nothing it says later is reported. A microphone opened for the prompt is released as soon as it
+ * opens, whether or not the press is still there.
+ *
+ * Errors. Opening the microphone for the prompt: refused (`NotAllowedError`, `SecurityError`) is `denied`; no
+ * microphone (`NotFoundError`) is `unsupported`; anything else is `failed`. The recognizer's
+ * (`SpeechRecognitionErrorEvent.error`): `not-allowed` or `service-not-allowed` while the microphone is confirmed for
+ * the press (the browser reports it granted, or opening it just worked) is tried once more after
+ * {@link BROWSER_GRANT_RETRY_MS}, quietly (Android refuses the first start after a fresh grant and takes the next), and
+ * after that it is `unsupported`: the speech service refusing, not a permission the visitor can give. Unconfirmed,
+ * `not-allowed` is `denied` (and the next tap asks the browser again) and `service-not-allowed` (iOS with Siri and
+ * Dictation off, Edge's policy) is `unsupported`; `audio-capture` (no microphone) is `unsupported`;
+ * `language-not-supported` tries the fallback language once, then `unsupported`; `network` is `failed`, and
+ * `unsupported` the second time running on the page; `no-speech` is a quiet spell (a new recognizer); anything else is
+ * `failed`. A failure once words were heard, whatever it is, gives those words instead (`done`).
+ */
+
+/** One alternative of a recognition result: what the browser heard. */
+interface RecognitionAlternativeLike {
+    readonly transcript: string;
+}
+/** One recognition result: final, or still being recognised. */
+interface RecognitionResultLike {
+    readonly isFinal: boolean;
+    readonly length: number;
+    readonly [index: number]: RecognitionAlternativeLike | undefined;
+}
+/** The whole result list of a recognizer, as each `result` event carries it. */
+interface RecognitionResultListLike {
+    readonly length: number;
+    readonly [index: number]: RecognitionResultLike | undefined;
+}
+/** What this module uses of a browser's `SpeechRecognition` (declared here: TypeScript's DOM library has none). */
+interface RecognitionLike {
+    lang: string;
+    continuous: boolean;
+    interimResults: boolean;
+    maxAlternatives: number;
+    onstart: (() => void) | null;
+    onresult: ((event: {
+        readonly results: RecognitionResultListLike;
+    }) => void) | null;
+    onerror: ((event: {
+        readonly error: string;
+    }) => void) | null;
+    onend: (() => void) | null;
+    start(): void;
+    stop(): void;
+    abort(): void;
+}
+/** What a press needs from the browser. The browser's own by default; a test hands in fakes. */
+interface BrowserAvatarMicSeams {
+    /** A new recognizer, or null where the browser has none. */
+    createRecognition: () => RecognitionLike | null;
+    /** A monotonic clock, in milliseconds, for the cooldowns. */
+    now: () => number;
+    /** {@link primeBrowserMicrophone}: called inside the tap, where the microphone is not known to be granted. */
+    primeMicrophone: () => Promise<void> | null;
+    /** {@link watchBrowserMicPermission}: called once, as the mic is made. */
+    watchMicPermission: (onState: (state: PermissionState) => void) => void;
+    /** {@link hasCoarsePointer}: read at each tap. */
+    touch: () => boolean;
+}
+/** The Avatar tab's mic through the browser's own recognizer. See the file's comment for what a press does. */
+declare function createBrowserAvatarMic(options: BrowserAvatarMicOptions, seams?: Partial<BrowserAvatarMicSeams>): AvatarMic;
+
+/** What the worklet posts for each frame. */
+interface CaptureFrameMessage {
+    /** s16le mono 16 kHz, `CAPTURE_FRAME_SAMPLES` samples. */
+    pcm: ArrayBuffer;
+    /** Microphone level over the frame, dBFS. */
+    rmsDb: number;
+    /** Reference-bus level over the same samples, dBFS (about -90 when silent, as it always is here). */
+    refDb: number;
+    /** Zero crossings per sample. */
+    zcr: number;
+    /** AudioContext time at the end of the frame, seconds. */
+    ctxTime: number;
+    seq: number;
+}
+
+/**
+ * BOFF-9797 (Stage C). The Avatar tab's own microphone capture: `getUserMedia`, through the capture worklet
+ * (`capture-worklet.ts`), into 16 kHz PCM frames.
+ *
+ * **Ported, not shared.** From `src/shared/assistant/voice/handsfree/micCapture.ts` in Algoshred/fe-libs at the
+ * commit `capture-worklet.ts` names (`CAPTURE_WORKLET_PORTED_FROM`; blob `8486c630`), with one difference: fe-libs
+ * routes everything its assistant plays into the worklet's second input (its playback reference, for barge-in), and the
+ * website plays nothing through this context (the avatar's voice is the room's own audio element, and the mic is
+ * offered only while no session plays), so no reference bus is made and the worklet has the microphone as its only
+ * input. `looksLikeHeadset`, which only fe-libs' barge-in reads, is not ported. `capture-worklet.test.ts` holds the
+ * parity tests.
+ *
+ * It opens `getUserMedia` with echo cancellation, noise suppression and auto gain on. Constraints are plain `true`,
+ * never `{ exact }`: `exact` throws OverconstrainedError on handsets that cannot honour one. The settings the browser
+ * actually applied are read back (`track.getSettings()`), because some Android handsets report `true` and apply nothing.
+ *
+ * `stop()` stops the track (the browser's recording indicator goes), disconnects the graph and closes the context. The
+ * caller does that the moment a press ends, whichever way it ends; `pagehide` does it too, whatever the caller does.
+ */
+
+interface MicCaptureHandle {
+    context: AudioContext;
+    micSource: MediaStreamAudioSourceNode;
+    track: MediaStreamTrack;
+    /** What the browser reports it applied. */
+    settings: MediaTrackSettings;
+    /** Stop everything and release the microphone. Safe to call more than once. */
+    stop(): void;
+}
+
+/** What the mic needs of a WebSocket; {@link openBrowserSocket} adapts the browser's. (fe-libs' seam.) */
+interface SocketLike {
+    readonly readyState: number;
+    send(data: string | ArrayBufferLike | ArrayBufferView): void;
+    close(code?: number): void;
+    listen(handlers: SocketHandlers): void;
+}
+interface SocketHandlers {
+    open(): void;
+    message(data: unknown): void;
+    close(code: number): void;
+    error(): void;
 }
 
 /**
@@ -176,7 +300,8 @@ interface AvatarMicModule {
  *
  * **Released on every way out.** The microphone's track is stopped the moment the recording stops, and also when the
  * press fails, is cancelled (`cancel()`: the view going, the page hidden, the Chat tab) or superseded (a new press while
- * one is still finishing: `cancel` for the old one, so it stops being paid for, and its text is never reported). A
+ * one is still finishing, a tap the view lets through for this stream mic only: `cancel` for the old one, so it stops
+ * being paid for, and its text is never reported). A
  * microphone that opens after its press has gone is closed as it opens. The stream is closed whenever no press needs
  * it: one connection serves a press and the press that supersedes it, and no other.
  */
@@ -190,4 +315,4 @@ interface AvatarMicSeams {
 /** The mic of one Avatar tab. See the file's comment for what a press does. */
 declare function createAvatarMic(options: AvatarMicOptions, seams?: Partial<AvatarMicSeams>): AvatarMic;
 
-export { type AvatarMic, type AvatarMicFailure, type AvatarMicModule, type AvatarMicOffer, type AvatarMicOptions, type AvatarMicState, createAvatarMic };
+export { type AvatarMic, type AvatarMicFailure, type AvatarMicModule, type AvatarMicOffer, type AvatarMicOptions, type AvatarMicState, type BrowserAvatarMicOptions, createAvatarMic, createBrowserAvatarMic };