feat(web): voice input — record, check the recognised text, send

Microphone button in the chat composer and in free-text steps. The browser
records 16 kHz mono WAV (no server-side ffmpeg), /me/{token}/stt transcribes,
the text lands in the field for the patient to check and edit — never sent
automatically, so safety and scrub() apply on send. Hidden when the server has
voice input off or the browser lacks a microphone API. nginx: 4 MB body limit.

Refs #117

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

web/nginx.conf

@@ -14,5 +14,6 @@ server {
         proxy_http_version 1.1;
         proxy_buffering off;
         proxy_read_timeout 120s;
+        client_max_body_size 4m; # voice input: up to 60 s of 16 kHz WAV (~1.9 MB)
     }
 }

web/src/api/client.ts

@@ -46,7 +46,23 @@ export const api = {
   signInAnonymous: (focus?: Focus) => post<AuthResult>("/auth/anonymous", { focus }),
   // Presentation entry: a fresh private copy of the demo patient, straight to the chat.
   signInDemoFresh: () => post<AuthResult>("/auth/demo/fresh", {}),
-  patientCapabilities: () => request<{ primary_screening: boolean }>("/patient/capabilities"),
+  patientCapabilities: () =>
+    request<{ primary_screening: boolean; voice_input?: boolean }>("/patient/capabilities"),
+  // Voice → text for the patient to check (#117). Raw WAV body; the server only transcribes.
+  transcribe: async (token: string, wav: Blob): Promise<string> => {
+    let res: Response;
+    try {
+      res = await fetch(`${BASE}/me/${encodeURIComponent(token)}/stt`, {
+        method: "POST",
+        headers: { "Content-Type": "audio/wav" },
+        body: wav,
+      });
+    } catch {
+      throw new ApiError(0, "network");
+    }
+    if (!res.ok) throw new ApiError(res.status, res.statusText);
+    return ((await res.json()) as { text: string }).text;
+  },
   connectDoctor: (token: string) => post<{ has_doctor: boolean }>(`/me/${encodeURIComponent(token)}/doctor/connect`),
   demoPersonas: () => request<DemoPersona[]>("/patient/demo-personas"),
   crisisResources: () => request<CrisisResource[]>("/crisis/resources"),

web/src/components/Icon.tsx

@@ -32,6 +32,12 @@ const paths = {
   arrowLeft: <path d="M19 12H5m6-6l-6 6 6 6" />,
   close: <path d="M6 6l12 12M18 6L6 18" />,
   check: <path d="M5 12.5l4.5 4.5L19 7.5" />,
+  mic: (
+    <>
+      <rect x="9" y="3" width="6" height="11" rx="3" />
+      <path d="M5.5 11a6.5 6.5 0 0 0 13 0M12 17.5V21" />
+    </>
+  ),
   x: <path d="M7 7l10 10M17 7L7 17" />,
   info: (
     <>

web/src/components/VoiceInput.tsx

@@ -0,0 +1,165 @@
+import { useEffect, useRef, useState } from "react";
+
+import { ApiError, api } from "../api/client";
+import { ru } from "../content/ru";
+import { useSession } from "../session/SessionContext";
+import { MAX_SECONDS, Recorder, voiceSupported } from "../voice/recorder";
+import { Icon } from "./Icon";
+import { useToast } from "./Toast";
+
+/*
+ * Voice input next to a free-text field (#117, slide 11: «Записать → Проверить»). The recognised text is put into
+ * the field for the patient to check and edit — never sent automatically: safety and scrub() run when they send it.
+ */
+
+let capability: Promise<boolean> | null = null;
+function voiceEnabled(): Promise<boolean> {
+  capability ??= api
+    .patientCapabilities()
+    .then((c) => Boolean(c.voice_input) && voiceSupported())
+    .catch(() => false);
+  return capability;
+}
+
+const BARS = 14;
+type State = "idle" | "recording" | "transcribing";
+
+export function VoiceInput({
+  disabled,
+  onText,
+  compact,
+}: {
+  disabled?: boolean;
+  onText: (text: string) => void;
+  /** Chat composer: a round icon button; the recording panel floats above the composer. */
+  compact?: boolean;
+}) {
+  const { session } = useSession();
+  const toast = useToast();
+  const [enabled, setEnabled] = useState(false);
+  const [state, setState] = useState<State>("idle");
+  const [seconds, setSeconds] = useState(0);
+  const [levels, setLevels] = useState<number[]>(() => Array(BARS).fill(0));
+  const rec = useRef<Recorder | null>(null);
+  const timer = useRef<number>();
+
+  useEffect(() => {
+    let alive = true;
+    void voiceEnabled().then((on) => alive && setEnabled(on));
+    return () => {
+      alive = false;
+      window.clearInterval(timer.current);
+      rec.current?.cancel();
+    };
+  }, []);
+
+  if (!enabled || !session) return null;
+
+  const start = async () => {
+    const r = new Recorder();
+    try {
+      await r.start((level) => setLevels((prev) => [...prev.slice(1), level]));
+    } catch {
+      r.cancel();
+      toast(ru.voice.noMic);
+      return;
+    }
+    rec.current = r;
+    setSeconds(0);
+    setState("recording");
+    const began = Date.now();
+    timer.current = window.setInterval(() => {
+      const s = Math.floor((Date.now() - began) / 1000);
+      setSeconds(s);
+      if (s >= MAX_SECONDS) void stop();
+    }, 250);
+  };
+
+  const reset = () => {
+    window.clearInterval(timer.current);
+    rec.current = null;
+    setLevels(Array(BARS).fill(0));
+  };
+
+  const stop = async () => {
+    const wav = rec.current?.stop() ?? null;
+    reset();
+    if (!wav) {
+      setState("idle");
+      return;
+    }
+    setState("transcribing");
+    try {
+      const text = await api.transcribe(session.token, wav);
+      onText(text);
+      toast(ru.voice.check);
+    } catch (e) {
+      const status = e instanceof ApiError ? e.status : 0;
+      toast(status === 422 ? ru.voice.noSpeech : status === 413 ? ru.voice.tooLong : ru.voice.failed);
+    } finally {
+      setState("idle");
+    }
+  };
+
+  const cancel = () => {
+    rec.current?.cancel();
+    reset();
+    setState("idle");
+  };
+
+  if (state === "idle" && compact) {
+    return (
+      <button
+        type="button"
+        className="mic"
+        disabled={disabled}
+        onClick={() => void start()}
+        aria-label={ru.voice.start}
+        title={ru.voice.start}
+      >
+        <Icon name="mic" size={20} />
+      </button>
+    );
+  }
+  if (state === "idle") {
+    return (
+      <button type="button" className="voice-btn" disabled={disabled} onClick={() => void start()}>
+        <Icon name="mic" size={20} />
+        {ru.voice.start}
+      </button>
+    );
+  }
+
+  const mm = String(Math.floor(seconds / 60)).padStart(2, "0");
+  const ss = String(seconds % 60).padStart(2, "0");
+  return (
+    <div className={compact ? "voice voice--float" : "voice"} role="group" aria-label={ru.voice.recording}>
+      {state === "recording" ? (
+        <>
+          <p className="voice__hint">{ru.voice.recording}</p>
+          <div className="voice__wave" aria-hidden="true">
+            {levels.map((l, i) => (
+              <span key={i} style={{ height: `${Math.max(12, Math.round(l * 100))}%` }} />
+            ))}
+          </div>
+          <div className="voice__row">
+            <span className="voice__time" aria-live="off">
+              {mm}:{ss}
+            </span>
+            <button type="button" className="btn btn--ghost voice__cancel" onClick={cancel}>
+              {ru.voice.cancel}
+            </button>
+            <button type="button" className="btn btn--primary voice__stop" onClick={() => void stop()}>
+              {ru.voice.stop}
+            </button>
+          </div>
+          <p className="voice__privacy">{ru.voice.privacy}</p>
+        </>
+      ) : (
+        <p className="voice__hint" aria-live="polite">
+          {ru.voice.transcribing}
+        </p>
+      )}
+    </div>
+  );
+}

web/src/content/ru.ts

@@ -251,6 +251,19 @@ export const ru = {
     drinksLead: "Посчитайте, сколько порций выходит в обычный день, когда вы выпиваете.",
     noticeTitle: "Опрос приостановлен",
   },
+  voice: {
+    start: "Сказать голосом",
+    stop: "Остановить запись",
+    cancel: "Отменить",
+    recording: "Говорите — запись идёт",
+    transcribing: "Распознаём речь…",
+    check: "Проверьте распознанный текст — его можно исправить перед отправкой.",
+    privacy: "Запись не сохраняется: её видит только система распознавания, текст — специалисты вашей программы.",
+    noMic: "Нет доступа к микрофону. Разрешите его в настройках браузера или напишите текстом.",
+    noSpeech: "Не удалось разобрать речь. Попробуйте ещё раз или напишите текстом.",
+    failed: "Распознавание сейчас недоступно. Напишите, пожалуйста, текстом.",
+    tooLong: "Запись слишком длинная — до минуты за раз.",
+  },
   messages: {
     typing: "Помощник печатает",
     title: "Сообщения",

web/src/screens/Flow.tsx

@@ -5,6 +5,7 @@ import type { Catalog, DialogButton } from "../api/types";
 import { AppShell, Notice, TopBar } from "../components/AppShell";
 import { HotlineCard, telHref, useHelp } from "../components/HelpSheet";
 import { useToast } from "../components/Toast";
+import { VoiceInput } from "../components/VoiceInput";
 import { Icon, type IconName } from "../components/Icon";
 import { useCatalog } from "../content/reference";
 import { ru } from "../content/ru";
@@ -259,6 +260,7 @@ function FreeText({
         />
         <span className="textarea__count">{text.length}/500</span>
       </div>
+      <VoiceInput disabled={busy} onText={(t) => setText((prev) => (prev.trim() ? `${prev.trim()} ${t}` : t).slice(0, 500))} />
       <div className="btn-row">
         {skip ? (
           <button type="button" className="btn btn--ghost" disabled={busy} onClick={() => onSkip(skip)}>

web/src/screens/Messages.tsx

@@ -3,6 +3,7 @@ import { useNavigate } from "react-router-dom";
 
 import { AppShell, TopBar } from "../components/AppShell";
 import { Icon } from "../components/Icon";
+import { VoiceInput } from "../components/VoiceInput";
 import { ru } from "../content/ru";
 import { useDialog } from "../dialog/DialogContext";
 
@@ -112,6 +113,11 @@ export function Messages() {
             onKeyDown={onKeyDown}
             maxLength={2000}
           />
+          <VoiceInput
+            compact
+            disabled={dialog.busy}
+            onText={(text) => setDraft((prev) => (prev.trim() ? `${prev.trim()} ${text}` : text).slice(0, 2000))}
+          />
           <button type="submit" className="send" disabled={dialog.busy || !draft.trim()} aria-label={ru.common.send}>
             <Icon name="send" size={20} />
           </button>

web/src/theme/global.css

@@ -1566,6 +1566,106 @@ button.row-card:hover:not(:disabled) {
   color: var(--c-text-3);
 }
 
+/* ---------------------------------------------------------------- voice input (#117, slide 11) */
+
+.voice-btn {
+  display: inline-flex;
+  align-items: center;
+  gap: var(--s-2);
+  justify-self: start;
+  min-height: 44px;
+  padding: 0 var(--s-4);
+  border: 1.5px solid var(--c-border-strong);
+  border-radius: 999px;
+  background: var(--c-surface);
+  color: var(--c-teal-strong);
+  font-weight: 600;
+  font-size: var(--fs-14);
+}
+
+.voice-btn:hover:not(:disabled) {
+  background: var(--c-teal-soft);
+}
+
+.mic {
+  flex: none;
+  width: 44px;
+  height: 44px;
+  display: grid;
+  place-items: center;
+  border: 1.5px solid var(--c-border);
+  border-radius: 50%;
+  background: var(--c-surface);
+  color: var(--c-teal-strong);
+}
+
+.mic:hover:not(:disabled) {
+  background: var(--c-teal-soft);
+}
+
+.voice {
+  display: grid;
+  gap: var(--s-3);
+  padding: var(--s-4);
+  border: 1px solid var(--c-border);
+  border-radius: var(--r-tile);
+  background: #fbf6ee; /* warm, as on the slide */
+}
+
+.voice--float {
+  position: absolute;
+  left: 0;
+  right: 0;
+  bottom: calc(100% + var(--s-2));
+  box-shadow: 0 8px 24px rgb(33 59 56 / 12%);
+}
+
+.voice__hint {
+  margin: 0;
+  font-weight: 600;
+  color: var(--c-text);
+}
+
+.voice__wave {
+  display: flex;
+  align-items: center;
+  gap: 5px;
+  height: 44px;
+}
+
+.voice__wave span {
+  flex: 1;
+  max-width: 6px;
+  border-radius: 3px;
+  background: #b5452f;
+  transition: height 90ms linear;
+}
+
+.voice__row {
+  display: grid;
+  grid-template-columns: auto 1fr 1.4fr;
+  align-items: center;
+  gap: var(--s-2);
+}
+
+.voice__time {
+  font-variant-numeric: tabular-nums;
+  font-weight: 600;
+  color: var(--c-text-2);
+}
+
+.voice__row .btn {
+  min-height: 44px;
+  padding: 0 var(--s-3);
+  font-size: var(--fs-14);
+}
+
+.voice__privacy {
+  margin: 0;
+  font-size: var(--fs-12);
+  color: var(--c-text-3);
+}
+
 /* ---------------------------------------------------------------- emergency */
 
 .sosblock {

web/src/voice/recorder.ts

@@ -0,0 +1,119 @@
+/*
+ * Voice recording for the patient's messages (#117, slide 11): microphone → 16 kHz mono 16-bit WAV in the browser,
+ * so the server needs no ffmpeg and Chrome, Android and iOS Safari produce the same format (ADR-014).
+ * The audio lives only in memory until it is sent for transcription; nothing is stored on the device.
+ */
+
+export const TARGET_RATE = 16_000;
+export const MAX_SECONDS = 60;
+
+type WindowWithWebkit = Window & { webkitAudioContext?: typeof AudioContext };
+
+export function voiceSupported(): boolean {
+  const w = window as WindowWithWebkit;
+  return typeof navigator.mediaDevices?.getUserMedia === "function" && Boolean(window.AudioContext || w.webkitAudioContext);
+}
+
+export class Recorder {
+  private ctx: AudioContext | null = null;
+  private stream: MediaStream | null = null;
+  private node: ScriptProcessorNode | null = null;
+  private chunks: Float32Array[] = [];
+  private rate = 48_000;
+
+  /** `onLevel` gets the loudness 0..1 of each audio block (for the waveform). */
+  async start(onLevel: (level: number) => void): Promise<void> {
+    this.stream = await navigator.mediaDevices.getUserMedia({
+      audio: { channelCount: 1, echoCancellation: true, noiseSuppression: true, autoGainControl: true },
+    });
+    const Ctx = window.AudioContext || (window as WindowWithWebkit).webkitAudioContext!;
+    this.ctx = new Ctx();
+    await this.ctx.resume();
+    this.rate = this.ctx.sampleRate;
+    const source = this.ctx.createMediaStreamSource(this.stream);
+    // ScriptProcessorNode is deprecated but works everywhere (incl. older iOS Safari) without a worklet file.
+    this.node = this.ctx.createScriptProcessor(4096, 1, 1);
+    this.node.onaudioprocess = (e) => {
+      const data = new Float32Array(e.inputBuffer.getChannelData(0));
+      this.chunks.push(data);
+      let sum = 0;
+      for (let i = 0; i < data.length; i++) sum += data[i] * data[i];
+      onLevel(Math.min(1, Math.sqrt(sum / data.length) * 4));
+    };
+    source.connect(this.node);
+    this.node.connect(this.ctx.destination);
+  }
+
+  /** Stops the microphone and returns the WAV, or null if nothing was recorded. */
+  stop(): Blob | null {
+    const chunks = this.chunks;
+    const rate = this.rate;
+    this.release();
+    if (!chunks.length) return null;
+    return encodeWav(downsample(concat(chunks), rate, TARGET_RATE), TARGET_RATE);
+  }
+
+  cancel(): void {
+    this.release();
+  }
+
+  private release(): void {
+    this.node?.disconnect();
+    if (this.node) this.node.onaudioprocess = null;
+    this.stream?.getTracks().forEach((t) => t.stop());
+    void this.ctx?.close();
+    this.node = null;
+    this.stream = null;
+    this.ctx = null;
+    this.chunks = [];
+  }
+}
+
+function concat(chunks: Float32Array[]): Float32Array {
+  const out = new Float32Array(chunks.reduce((n, c) => n + c.length, 0));
+  let pos = 0;
+  for (const c of chunks) {
+    out.set(c, pos);
+    pos += c.length;
+  }
+  return out;
+}
+
+/** Averaging downsampler: good enough for speech, avoids aliasing better than dropping samples. */
+function downsample(input: Float32Array, from: number, to: number): Float32Array {
+  if (from === to) return input;
+  const ratio = from / to;
+  const out = new Float32Array(Math.floor(input.length / ratio));
+  for (let i = 0; i < out.length; i++) {
+    const start = Math.floor(i * ratio);
+    const end = Math.min(input.length, Math.floor((i + 1) * ratio));
+    let sum = 0;
+    for (let j = start; j < end; j++) sum += input[j];
+    out[i] = sum / Math.max(1, end - start);
+  }
+  return out;
+}
+
+function encodeWav(samples: Float32Array, rate: number): Blob {
+  const buf = new ArrayBuffer(44 + samples.length * 2);
+  const v = new DataView(buf);
+  const str = (o: number, s: string) => [...s].forEach((ch, i) => v.setUint8(o + i, ch.charCodeAt(0)));
+  str(0, "RIFF");
+  v.setUint32(4, 36 + samples.length * 2, true);
+  str(8, "WAVE");
+  str(12, "fmt ");
+  v.setUint32(16, 16, true);
+  v.setUint16(20, 1, true); // PCM
+  v.setUint16(22, 1, true); // mono
+  v.setUint32(24, rate, true);
+  v.setUint32(28, rate * 2, true); // byte rate
+  v.setUint16(32, 2, true); // block align
+  v.setUint16(34, 16, true);
+  str(36, "data");
+  v.setUint32(40, samples.length * 2, true);
+  for (let i = 0; i < samples.length; i++) {
+    const s = Math.max(-1, Math.min(1, samples[i]));
+    v.setInt16(44 + i * 2, s < 0 ? s * 0x8000 : s * 0x7fff, true);
+  }
+  return new Blob([buf], { type: "audio/wav" });
+}