Ask whether to transcribe a recording or send it as a voice note

e2e/USER_STORIES.md

@@ -944,10 +944,26 @@ Acceptance:
 
 - The dictation button with no OpenRouter key saved asks for one, the same
   prompt reading a message out loud uses.
-- Recording and stopping puts the transcript in the composer, ready to send.
+- Stopping a recording asks whether to transcribe it or send it as a voice
+  note.
+- Choosing to transcribe puts the transcript in the composer, ready to send.
 - Leaving the room while a transcription is still out does not lose it: the
   transcript lands in the composer that is there when it comes back.
 
+### US-126 — Send a voice note
+
+As alice, I want to send a recording as it is, so that the message carries my
+voice rather than a transcript of it.
+
+Acceptance:
+
+- Choosing to send a voice note uploads the recording and attaches it to the
+  composer.
+- Sending it publishes the message with the audio, which renders as a player in
+  the timeline.
+- Discarding the recording instead leaves the composer empty and uploads
+  nothing.
+
 ## Rich content & media rendering
 
 ### US-060 — Reveal a flagged sensitive message

e2e/specs/composer.spec.ts

@@ -425,6 +425,15 @@ const dictateButton = (page: Page) => page.getByRole("button", {name: "Start dic
 
 const stopButton = (page: Page) => page.getByRole("button", {name: "Stop recording"})
 
+// Every recording ends at the same question, so each spec starts from the answer it is about.
+const record = async (page: Page) => {
+  await dictateButton(page).click()
+  await expect(stopButton(page)).toBeVisible()
+  await stopButton(page).click()
+
+  return dialog(page, "Transcribe or send?")
+}
+
 test("US-125 dictate a message", async ({seed, as}) => {
   const scenario = await seed(({relay, user}) => {
     const space = relay("space")
@@ -440,8 +449,11 @@ test("US-125 dictate a message", async ({seed, as}) => {
   const alice = await as(users.alice, roomPath(url, "general"))
   const transcription = await mockOpenRouterTranscription(alice.context(), "the tide turns at six")
 
-  // With no key saved, dictation asks for one, the same prompt reading a message out loud uses.
+  // Recording asks for nothing. Asking for a transcript with no key saved asks for one, the same
-  await dictateButton(alice).click()
+  // prompt reading a message out loud uses.
+  const action = await record(alice)
+
+  await action.getByRole("button", {name: "Transcribe it"}).click()
 
   const enable = dialog(alice, "Enable voice input?")
 
@@ -450,10 +462,8 @@ test("US-125 dictate a message", async ({seed, as}) => {
 
   await expect(alice.getByRole("alert")).toContainText("Voice input is ready to use!")
 
-  await dictateButton(alice).click()
+  // The recording is still waiting behind that prompt for the answer it asked for.
-  await expect(stopButton(alice)).toBeVisible()
+  await action.getByRole("button", {name: "Transcribe it"}).click()
-
-  await stopButton(alice).click()
 
   await expect(composer(alice)).toContainText("the tide turns at six")
 
@@ -469,10 +479,7 @@ test("US-125 dictate a message", async ({seed, as}) => {
   // left while the request is still out, and the transcript waits for whichever composer is next.
   transcription.hold()
 
-  await dictateButton(alice).click()
+  await (await record(alice)).getByRole("button", {name: "Transcribe it"}).click()
-  await expect(stopButton(alice)).toBeVisible()
-
-  await stopButton(alice).click()
 
   // In-app rather than a fresh load: a dictation is held by the app rather than by the composer
   // that started one, so reloading the page is losing it rather than leaving it.
@@ -486,3 +493,35 @@ test("US-125 dictate a message", async ({seed, as}) => {
 
   await expect(composer(alice)).toContainText("the tide turns at six")
 })
+
+test("US-126 send a voice note", async ({seed, as}) => {
+  const scenario = await seed(({relay, user}) => {
+    const space = relay("space")
+
+    space.room("general", {name: "General"})
+    space.join(user.alice, "general")
+  })
+
+  const {url} = scenario.space("space")
+  const alice = await as(users.alice, roomPath(url, "general"))
+
+  await mockBlossom(alice.context(), {server: DEFAULT_BLOSSOM_ORIGIN})
+
+  // Leaving the question unanswered throws the recording away, so nothing is uploaded and the
+  // composer is where it was.
+  await (await record(alice)).getByRole("button", {name: "Discard"}).click()
+
+  await expect(dialog(alice, "Transcribe or send?")).toHaveCount(0)
+  await expect(composer(alice)).toHaveText("")
+
+  // Sending the recording as it is needs no OpenRouter key, only somewhere to upload it.
+  await (await record(alice)).getByRole("button", {name: "Send a voice note"}).click()
+
+  await expect(composer(alice)).toContainText(DEFAULT_BLOSSOM_ORIGIN)
+
+  await composer(alice).press("Enter")
+
+  // The imeta on the message says the upload is audio, which is what gives it a player rather
+  // than a link.
+  await expect(timeline(alice).locator(`audio[src^="${DEFAULT_BLOSSOM_ORIGIN}/"]`)).toBeVisible()
+})

src/app/components/ChatCompose.svelte

@@ -73,6 +73,14 @@
     ed.chain().focus().insertContent(escapeHtml(text)).run()
   }
 
+  const attachVoiceNote = async (audio: File) => {
+    const ed = await editor
+
+    ed.chain()
+      .addFile(audio, ed.state.selection.from + 1)
+      .run()
+  }
+
   const submit = async () => {
     if ($uploading || disabled) return
 
@@ -156,7 +164,11 @@
     <EditorContent {autofocus} {editor} />
   </div>
   {#if dictating || ($empty && !disabled)}
-    <DictationButton key={dictationKey} bind:dictating onTranscript={insertTranscript} />
+    <DictationButton
+      key={dictationKey}
+      bind:dictating
+      onTranscript={insertTranscript}
+      onVoiceNote={attachVoiceNote} />
   {:else}
     <Button
       data-tip="{window.navigator.platform.includes('Mac') ? 'cmd' : 'ctrl'}+enter to send"

src/app/components/DictationAction.svelte

@@ -0,0 +1,101 @@
+<script lang="ts">
+  import {onDestroy} from "svelte"
+  import DocumentText from "@assets/icons/document-text.svg?dataurl"
+  import Soundwave from "@assets/icons/soundwave.svg?dataurl"
+  import TrashBin from "@assets/icons/trash-bin-minimalistic.svg?dataurl"
+  import Icon from "@lib/components/Icon.svelte"
+  import Button from "@lib/components/Button.svelte"
+  import CardButton from "@lib/components/CardButton.svelte"
+  import Modal from "@lib/components/Modal.svelte"
+  import ModalBody from "@lib/components/ModalBody.svelte"
+  import ModalHeader from "@lib/components/ModalHeader.svelte"
+  import ModalTitle from "@lib/components/ModalTitle.svelte"
+  import ModalSubtitle from "@lib/components/ModalSubtitle.svelte"
+  import ModalFooter from "@lib/components/ModalFooter.svelte"
+  import OpenRouterEnable from "@app/components/OpenRouterEnable.svelte"
+  import {getSetting} from "@app/settings"
+  import {popModal, pushModal} from "@app/modal"
+
+  type Props = {
+    onTranscribe: () => void
+    onVoiceNote: () => void
+    onDiscard: () => void
+  }
+
+  const {onTranscribe, onVoiceNote, onDiscard}: Props = $props()
+
+  const transcribe = () => {
+    if (getSetting("openrouter_key")) {
+      pending = false
+      popModal()
+      onTranscribe()
+    } else {
+      // Nested, so that the recording is still here to transcribe once a key has been saved.
+      pushModal(
+        OpenRouterEnable,
+        {
+          feature: "Voice input",
+          subtitle: "Dictate your messages instead of typing them.",
+        },
+        {nested: true},
+      )
+    }
+  }
+
+  const sendVoiceNote = () => {
+    pending = false
+    popModal()
+    onVoiceNote()
+  }
+
+  let pending = true
+
+  // Leaving without choosing throws the recording away, whether that was the discard button, the
+  // escape key or a navigation out of the conversation.
+  onDestroy(() => {
+    if (pending) {
+      onDiscard()
+    }
+  })
+</script>
+
+<Modal>
+  <ModalBody>
+    <ModalHeader>
+      <ModalTitle>Transcribe or send?</ModalTitle>
+      <ModalSubtitle>Your recording can become text, or go as it is.</ModalSubtitle>
+    </ModalHeader>
+    <Button onclick={transcribe}>
+      <CardButton primary>
+        {#snippet icon()}
+          <div><Icon icon={DocumentText} size={7} /></div>
+        {/snippet}
+        {#snippet title()}
+          <div>Transcribe it</div>
+        {/snippet}
+        {#snippet info()}
+          <div>Turn what you said into text you can edit before sending.</div>
+        {/snippet}
+      </CardButton>
+    </Button>
+    <Button onclick={sendVoiceNote}>
+      <CardButton>
+        {#snippet icon()}
+          <div><Icon icon={Soundwave} size={7} /></div>
+        {/snippet}
+        {#snippet title()}
+          <div>Send a voice note</div>
+        {/snippet}
+        {#snippet info()}
+          <div>Attach the recording itself to your message.</div>
+        {/snippet}
+      </CardButton>
+    </Button>
+  </ModalBody>
+  <ModalFooter>
+    <Button class="button button-link" onclick={popModal}>
+      <Icon icon={TrashBin} />
+      Discard
+    </Button>
+  </ModalFooter>
+</Modal>

src/app/components/DictationButton.svelte

@@ -8,9 +8,8 @@
   import Button from "@lib/components/Button.svelte"
   import Spinner from "@lib/components/Spinner.svelte"
   import {errorMessage} from "@lib/util"
-  import OpenRouterEnable from "@app/components/OpenRouterEnable.svelte"
+  import DictationAction from "@app/components/DictationAction.svelte"
-  import {clearDictation, getDictation, startDictation} from "@app/dictation"
+  import {clearDictation, getDictation, startDictation, transcribeDictation} from "@app/dictation"
-  import {getSetting} from "@app/settings"
   import {pushModal} from "@app/modal"
   import {pushToast} from "@app/toast"
 
@@ -18,9 +17,10 @@
     key: string
     dictating?: boolean
     onTranscript: (text: string) => MaybeAsync<void>
+    onVoiceNote: (audio: File) => MaybeAsync<void>
   }
 
-  let {key, dictating = $bindable(false), onTranscript}: Props = $props()
+  let {key, dictating = $bindable(false), onTranscript, onVoiceNote}: Props = $props()
 
   // Room noise idles just below this, so the pulse follows quiet speech too.
   const onLevel = (level: number) => {
@@ -28,7 +28,6 @@
   }
 
   const start = async () => {
-    if (getSetting("openrouter_key")) {
     // Granting microphone access can sit on a permission prompt for a while, so hold the button
     // until it resolves — a second click would open a stream nothing is left holding on to.
     loading = true
@@ -42,12 +41,6 @@
     } finally {
       loading = false
     }
-    } else {
-      pushModal(OpenRouterEnable, {
-        feature: "Voice input",
-        subtitle: "Dictate your messages instead of typing them.",
-      })
-    }
   }
 
   const stop = () => {
@@ -55,14 +48,52 @@
 
     recording = false
     loud = false
+    // Hold the button until the recording has been spent, rather than offering a second one
+    // behind the modal.
+    loading = true
 
+    choose()
+  }
+
+  const choose = () =>
+    pushModal(DictationAction, {
+      onTranscribe: transcribe,
+      onVoiceNote: attach,
+      onDiscard: discard,
+    })
+
+  const transcribe = () => {
+    const dictation = getDictation(key)
+
+    if (dictation) {
+      transcribeDictation(dictation)
       deliver()
     }
+  }
+
+  const attach = async () => {
+    const dictation = getDictation(key)
+
+    if (dictation) {
+      const audio = await dictation.audio
+
+      clearDictation(key)
+      loading = false
+
+      await onVoiceNote(audio)
+    }
+  }
+
+  const discard = () => {
+    clearDictation(key)
+
+    loading = false
+  }
 
   const deliver = async () => {
     const dictation = getDictation(key)
 
-    if (dictation) {
+    if (dictation?.finished) {
       loading = true
 
       await dictation.finished
@@ -92,7 +123,7 @@
   let destroyed = false
   let recording = $state(false)
   // Starts out in flight when a dictation is already waiting to be picked up, so that the composer
-  // keeps rendering this button until its transcript has been handed over.
+  // keeps rendering this button until the recording has been spent.
   let loading = $state(Boolean(getDictation(key)))
   let loud = $state(false)
 
@@ -108,8 +139,17 @@
     dictating = recording || loading
   })
 
-  // Pick up a dictation an earlier composer left running.
+  // Pick up a dictation an earlier composer left running, either where it was already transcribing
-  onMount(deliver)
+  // or at the choice the speaker never got to make.
+  onMount(() => {
+    const dictation = getDictation(key)
+
+    if (dictation?.finished) {
+      deliver()
+    } else if (dictation) {
+      choose()
+    }
+  })
 
   onDestroy(() => {
     destroyed = true
@@ -124,7 +164,7 @@
 
 <Button
   class={buttonClass}
-  data-tip={recording ? "Stop and transcribe" : "Record a message"}
+  data-tip={recording ? "Stop recording" : "Record a message"}
   aria-label={recording ? "Stop recording" : "Start dictation"}
   disabled={loading}
   onclick={toggle}>

src/app/components/RoomCompose.svelte

@@ -76,6 +76,14 @@
     ed.chain().focus().insertContent(escapeHtml(transcript)).run()
   }
 
+  const attachVoiceNote = async (audio: File) => {
+    const ed = await editor
+
+    ed.chain()
+      .addFile(audio, ed.state.selection.from + 1)
+      .run()
+  }
+
   // Argument tokens are whitespace-delimited, so separate one from whatever precedes it.
   const insertCommandToken = async (token: string) => {
     const ed = await editor
@@ -183,7 +191,11 @@
     <EditorContent {autofocus} {editor} />
   </div>
   {#if dictating || $empty}
-    <DictationButton {key} bind:dictating onTranscript={insertTranscript} />
+    <DictationButton
+      {key}
+      bind:dictating
+      onTranscript={insertTranscript}
+      onVoiceNote={attachVoiceNote} />
   {:else}
     <Button
       data-tip="{window.navigator.platform.includes('Mac') ? 'cmd' : 'ctrl'}+enter to send"

src/app/dictation.ts

@@ -12,12 +12,11 @@ const EXTENSIONS_BY_MIME_TYPE: Record<string, string> = {
   "audio/wav": "wav",
 }
 
-const transcribe = async (audio: Blob) => {
+const transcribe = async (audio: File) => {
-  const [mimeType] = audio.type.split(";")
   const body = new FormData()
 
   body.append("model", TRANSCRIPTION_MODEL)
-  body.append("file", audio, `dictation.${EXTENSIONS_BY_MIME_TYPE[mimeType] || "webm"}`)
+  body.append("file", audio, audio.name)
 
   const response = await fetch("https://openrouter.ai/api/v1/audio/transcriptions", {
     method: "POST",
@@ -37,9 +36,12 @@ const transcribe = async (audio: Blob) => {
 export type Dictation = {
   recording: boolean
   stop: () => void
-  // Resolves once the transcript or the error is on the dictation, so that awaiting it never takes
+  // Resolves once recording has stopped, with the audio named for the format the recorder chose.
-  // the result out of the registry — whoever is still around when it lands reads it from there.
+  audio: Promise<File>
-  finished: Promise<void>
+  // Set when the speaker asks for a transcript, and resolves once that or the error is on the
+  // dictation, so that awaiting it never takes the result out of the registry — whoever is still
+  // around when it lands reads it from there.
+  finished?: Promise<void>
   transcript?: string
   error?: unknown
 }
@@ -53,6 +55,17 @@ export const getDictation = (key: string) => dictations.get(key)
 
 export const clearDictation = (key: string) => dictations.delete(key)
 
+export const transcribeDictation = (dictation: Dictation) => {
+  dictation.finished = dictation.audio.then(transcribe).then(
+    transcript => {
+      dictation.transcript = transcript
+    },
+    error => {
+      dictation.error = error
+    },
+  )
+}
+
 // Reports how loud the microphone is once per frame, so the caller can show the speaker that we're
 // hearing them. Levels follow the waveform's envelope — jumping to each peak, then decaying — since
 // the raw root mean square drops to nothing in the gaps between words.
@@ -93,7 +106,7 @@ export const startDictation = async (key: string, onLevel: (level: number) => vo
 
   let frame = requestAnimationFrame(measure)
 
-  const audio = new Promise<Blob>(resolve => {
+  const audio = new Promise<File>(resolve => {
     recorder.addEventListener("stop", () => {
       cancelAnimationFrame(frame)
       context.close()
@@ -102,7 +115,12 @@ export const startDictation = async (key: string, onLevel: (level: number) => vo
         track.stop()
       }
 
-      resolve(new Blob(chunks, {type: recorder.mimeType}))
+      // The recorder names its codec alongside the container, which is more than the imeta on a
+      // voice note or the extension on a blossom url can carry, so keep the container alone.
+      const [type] = recorder.mimeType.split(";")
+      const extension = EXTENSIONS_BY_MIME_TYPE[type] || "webm"
+
+      resolve(new File(chunks, `dictation.${extension}`, {type}))
     })
   })
 
@@ -112,14 +130,7 @@ export const startDictation = async (key: string, onLevel: (level: number) => vo
       dictation.recording = false
       recorder.stop()
     },
-    finished: audio.then(transcribe).then(
+    audio,
-      transcript => {
-        dictation.transcript = transcript
-      },
-      error => {
-        dictation.error = error
-      },
-    ),
   }
 
   dictations.set(key, dictation)

src/routes/settings/content/+page.svelte

@@ -159,8 +159,8 @@
         {#snippet info()}
           <p>
             Add an <Link external href="https://openrouter.ai/settings/keys" class="text-primary"
-              >OpenRouter API key</Link> to dictate messages using the microphone button in your composer,
+              >OpenRouter API key</Link> to transcribe what you record with the microphone button in your
-            and to have messages read out loud to you.
+            composer, and to have messages read out loud to you.
           </p>
         {/snippet}
       </Field>