@dxos/plugin-transcription 0.8.3 → 0.8.4-main.03d5cd7b56

package/src/normalization/normalization.ts

@@ -0,0 +1,76 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import * as Effect from 'effect/Effect';
+import * as Schema from 'effect/Schema';
+
+import { Operation } from '@dxos/compute';
+import { log } from '@dxos/log';
+import { trim } from '@dxos/util';
+
+import { TranscriptOperation } from '../types';
+
+type MessageWithRangeIdType = TranscriptOperation.MessageWithRangeIdType;
+type SentenceNormalizationInputType = TranscriptOperation.SentenceNormalizationInputType;
+const { SentenceNormalization, SentenceNormalizationOutput } = TranscriptOperation;
+
+export type MessageWithRangeId = MessageWithRangeIdType;
+
+/** @deprecated Use SentenceNormalizationInputType */
+export type NormalizationInput = SentenceNormalizationInputType;
+
+/** @deprecated Use SentenceNormalizationOutput */
+export type NormalizationOutput = Schema.Schema.Type<typeof SentenceNormalizationOutput>;
+
+/**
+ * Sentence normalization for transcription.
+ * TODO(dmaretskyi): Reimplement using AiRequest.run or LanguageModel.generateObject - runStructured was removed.
+ */
+export const sentenceNormalization = SentenceNormalization.pipe(
+  Operation.withHandler(
+    Effect.fn(function* ({ messages }) {
+      log.info('input', { messages });
+      // TODO(dmaretskyi): runStructured was removed from AiRequest. Reimplement using new API.
+      return yield* Effect.die(
+        new Error(
+          'Sentence normalization needs to be reimplemented - runStructured was removed from AiRequest. Use AiRequest.run or LanguageModel.generateObject.',
+        ),
+      );
+    }),
+  ),
+);
+
+const _unused = trim`
+  You are observing a real-time transcript of a single person speaking.
+  The transcription is delivered in chunks of 10 seconds or less. As a result, individual sentences may be split across multiple messages, or multiple sentences may appear within a single message. Additionally, because this is real-time transcription, punctuation and capitalization may be incorrect or missing.
+
+  # Task Description:
+    - Your task is to detect and reconstruct broken or incomplete sentences by merging segments into coherent, grammatically correct sentences where appropriate.
+
+  # Input Format:
+    - The input is an array of messages that contain transcription blocks.
+    - Each message is a Message.Message, as described in the output format.
+
+  # Output Format:
+    - You have been provided with the tool that defines the output format; make sure to query it.
+    - Do not output anything other than the expected format.
+
+  # Segment Handling Rules:
+    - Sort messages by timestamp.
+    - Leave ID of the first message in a sentence.
+    - Each divided sentence should be added to the output as a separate message or block within existing messages, preserving the original order.
+    - If one message contains multiple sentences, you should not split them.
+    - The last sentence may be incomplete; it should still be output as a separate block or message.
+    - Keep track of the original timestamps of the messages and blocks and, use the timestamp of the first message/block of a sentence as the 'started' field of the merged message.
+    - The 'rangeId' field of the merged message should include 'messageId'-s and 'rangeId'-s of all messages that have been used to construct this message.
+
+  # Punctuation and Capitalization:
+    - Do not rely on the original punctuation and capitalization—they may be incorrect.
+    - Use logical reasoning to apply appropriate punctuation (e.g., period, comma, question mark, exclamation mark) and capitalization.
+
+  # Restrictions:
+    - Do not alter the order of the original messages; maintain the natural flow of speech.
+    - Do not interpret or infer meaning beyond reconstructing sentences.
+    - Do not add or remove any words or phrases.
+`;

package/src/{segments-normalization → normalization}/utils.ts

@@ -2,8 +2,8 @@
 // Copyright 2025 DXOS.org
 //
 
-import { type DataType } from '@dxos/schema';
+import { type Actor } from '@dxos/types';
 
-export const getActorId = (actor: DataType.Actor) => {
+export const getActorId = (actor: Actor.Actor) => {
   return actor.identityDid || actor.identityKey || actor.email || actor.name || '';
 };

package/src/operations/create.ts

@@ -0,0 +1,24 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import * as Effect from 'effect/Effect';
+
+import { Operation } from '@dxos/compute';
+import { Feed, Ref } from '@dxos/echo';
+import { Transcript } from '@dxos/types';
+
+import { TranscriptOperation } from '../types';
+
+const handler: Operation.WithHandler<typeof TranscriptOperation.Create> = TranscriptOperation.Create.pipe(
+  Operation.withHandler(
+    Effect.fn(function* ({ space }) {
+      const feed = space.db.add(Feed.make());
+      return {
+        object: Transcript.make(Ref.make(feed)),
+      };
+    }),
+  ),
+);
+
+export default handler;

package/src/operations/index.ts

@@ -0,0 +1,11 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import { OperationHandlerSet } from '@dxos/compute';
+
+export const TranscriptionOperationHandlerSet = OperationHandlerSet.lazy(
+  () => import('./create'),
+  () => import('./open'),
+  () => import('./summarize'),
+);

package/src/operations/open.ts

@@ -0,0 +1,28 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import * as Effect from 'effect/Effect';
+
+import { Operation } from '@dxos/compute';
+import { Database, Feed, Filter } from '@dxos/echo';
+import { Message } from '@dxos/types';
+
+import { TranscriptOperation } from '../types';
+import { renderByline } from '../util';
+
+const handler: Operation.WithHandler<typeof TranscriptOperation.Open> = TranscriptOperation.Open.pipe(
+  Operation.withHandler(
+    Effect.fn(function* ({ transcript }) {
+      const transcriptObj = yield* Database.load(transcript);
+      const feed = yield* Database.load(transcriptObj.feed);
+      const messages = yield* Feed.runQuery(feed, Filter.type(Message.Message));
+      const content = messages
+        .flatMap((message: Message.Message, index: number) => renderByline([])(message, index))
+        .join('\n\n');
+      return { content };
+    }),
+  ),
+);
+
+export default handler;

package/src/operations/summarize.ts

@@ -0,0 +1,101 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import * as Array from 'effect/Array';
+import * as Effect from 'effect/Effect';
+import * as Function from 'effect/Function';
+import * as Layer from 'effect/Layer';
+import * as Option from 'effect/Option';
+
+import { AiService, ConsolePrinter, ToolExecutionService, ToolResolverService } from '@dxos/ai';
+import { AiRequest, GenerationObserver } from '@dxos/assistant';
+import { Trace, Operation, OperationRegistry } from '@dxos/compute';
+import { Database } from '@dxos/echo';
+import { trim } from '@dxos/util';
+
+import { TranscriptOperation } from '../types';
+
+/**
+ * Summarize a transcript of a meeting.
+ */
+const handler: Operation.WithHandler<typeof TranscriptOperation.Summarize> = TranscriptOperation.Summarize.pipe(
+  Operation.withHandler(
+    Effect.fnUntraced(
+      function* ({ transcript, notes }) {
+        const result = yield* new AiRequest({
+          observer: GenerationObserver.fromPrinter(new ConsolePrinter({ tag: 'summarize' })),
+        }).run({
+          prompt: `Transcript: ${transcript}\n\nNotes: ${notes}`,
+          history: [],
+          system: systemPrompt,
+        });
+
+        const summary = Function.pipe(
+          result,
+          Array.findLast((msg) => msg.sender.role === 'assistant' && msg.blocks.some((block) => block._tag === 'text')),
+          Option.flatMap((msg) =>
+            Function.pipe(
+              msg.blocks,
+              Array.findLast((block) => block._tag === 'text'),
+              Option.map((block) => block.text),
+            ),
+          ),
+          Option.getOrThrowWith(() => new Error('No summary found')),
+        );
+
+        return { summary };
+      },
+      Effect.provide(
+        Layer.mergeAll(
+          AiService.model('@anthropic/claude-sonnet-4-0'),
+          ToolResolverService.layerEmpty,
+          ToolExecutionService.layerEmpty,
+          Trace.writerLayerNoop,
+          Database.notAvailable,
+          Layer.succeed(Operation.Service, {
+            invoke: () => Effect.die('Not available.'),
+            schedule: () => Effect.die('Not available.'),
+            invokePromise: async () => ({ error: new Error('Not available.') }),
+          } as any),
+          Layer.succeed(OperationRegistry.Service, { resolve: () => Effect.succeed(undefined) } as any),
+        ),
+      ),
+    ),
+  ),
+);
+
+export default handler;
+
+const systemPrompt = trim`
+  You are a helpful assistant that summarizes transcripts of meetings.
+
+  # Goal
+  Create a markdown summary of the meeting transcript with text notes provided.
+  Notes are very important so make sure to include them in the summary if they contain meaningful information.
+
+  # Formatting
+  - Format the summary as a markdown document without extra comments like "Here is the summary of the transcript:".
+  - Use markdown formatting for headings and bullet points.
+  - Format the summary as a list of key points and takeaways.
+  - All names of people should be in bold.
+
+  # Note Taking
+  - Correlate items in the summary with the person of origin to build a coherent narrative.
+  - Include short quotes verbatim where appropriate. Especially when concerned with design decisions and problem descriptions.
+
+  # Tasks
+  At the end of the summary include tasks.
+  Extract only the tasks that are:
+  - Directly actionable.
+  - Clearly assigned to a person or team (or can easily be inferred).
+  - Strongly implied by the conversation and/or user note (no speculative tasks).
+  - Specific enough that someone reading them would know exactly what to do next.
+
+  Format all tasks as markdown checkboxes using the syntax:
+  - [ ] Task description.
+
+  Additional information can be included (indented).
+
+  If no actionable tasks are found, omit this tasks section.
+`;

package/src/plugin.ts

@@ -0,0 +1,11 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import { Plugin } from '@dxos/app-framework';
+
+import { meta } from './meta';
+
+export const TranscriptionPlugin = Plugin.lazy(meta, () => import('#plugin'));
+
+export { TranscriptionOperationHandlerSet } from './operations';

package/src/stories/FileTranscription.stories.tsx

@@ -0,0 +1,269 @@
+//
+// Copyright 2025 DXOS.org
+//
+
+import { type Meta, type StoryObj } from '@storybook/react-vite';
+import * as Effect from 'effect/Effect';
+import React, { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+
+import { withPluginManager } from '@dxos/app-framework/testing';
+import { AppActivationEvents } from '@dxos/app-toolkit';
+import { scheduleTask } from '@dxos/async';
+import { Context } from '@dxos/context';
+import { Obj } from '@dxos/echo';
+import { MemoryQueue } from '@dxos/echo-db';
+import { createQueueDXN } from '@dxos/echo/internal';
+import { FunctionExecutor, ServiceContainer } from '@dxos/functions-runtime';
+import { log } from '@dxos/log';
+import { ClientPlugin } from '@dxos/plugin-client/plugin';
+import { initializeIdentity } from '@dxos/plugin-client/testing';
+import { PreviewPlugin } from '@dxos/plugin-preview/testing';
+import { StorybookPlugin, corePlugins } from '@dxos/plugin-testing';
+import { withLayout } from '@dxos/react-ui/testing';
+import { TestSchema } from '@dxos/schema/testing';
+import { type Actor, Message, Organization, Person } from '@dxos/types';
+import { seedTestData } from '@dxos/types/testing';
+
+import { useAudioFile, useQueueModelAdapter, useTranscriber } from '#hooks';
+import { TestItem } from '#testing';
+
+import { MessageNormalizer, getActorId } from '../normalization';
+import { type MediaStreamRecorderProps, type TranscriberProps } from '../transcriber';
+import { TranscriptionPlugin } from '../TranscriptionPlugin';
+import { renderByline } from '../util';
+import { TranscriptionStory } from './TranscriptionStory';
+import { useIsSpeaking } from './useIsSpeaking';
+
+const AudioFile = ({
+  detectSpeaking,
+  audioUrl,
+  normalizeSentences,
+  transcriberConfig,
+  recorderConfig,
+  audioConstraints,
+}: {
+  detectSpeaking?: boolean;
+  normalizeSentences?: boolean;
+  audioUrl: string;
+  transcriberConfig?: TranscriberProps['config'];
+  recorderConfig?: MediaStreamRecorderProps['config'];
+  audioConstraints?: MediaTrackConstraints;
+}) => {
+  const [running, setRunning] = useState(false);
+  const actor: Actor.Actor = useMemo(() => ({ name: 'You' }), []);
+
+  // Optional uploaded file overrides the default URL.
+  // The useEffect below revokes the previous URL whenever it changes (and on unmount).
+  const [uploadedUrl, setUploadedUrl] = useState<string>();
+  useEffect(() => {
+    return () => {
+      if (uploadedUrl) {
+        URL.revokeObjectURL(uploadedUrl);
+      }
+    };
+  }, [uploadedUrl]);
+
+  const handleUpload = useCallback((file: File) => {
+    setRunning(false);
+    setUploadedUrl(URL.createObjectURL(file));
+  }, []);
+
+  const sourceUrl = uploadedUrl ?? audioUrl;
+
+  // Audio.
+  const { audio, track, stream } = useAudioFile(sourceUrl, audioConstraints);
+
+  // Speaking.
+  const isSpeaking = detectSpeaking ? useIsSpeaking(track) : true;
+  const ref = useRef<HTMLAudioElement>(null);
+  useEffect(() => {
+    if (ref.current && stream) {
+      ref.current.srcObject = stream;
+    }
+  }, [stream]);
+
+  useEffect(() => {
+    if (!audio) {
+      return;
+    }
+
+    if (running) {
+      void audio.play();
+    } else {
+      void audio.pause();
+    }
+  }, [audio, running]);
+
+  // Transcriber.
+  // TODO(dmaretskyi): Use space.queues.create() instead.
+  const queueDxn = useMemo(() => createQueueDXN(), []);
+  const queue = useMemo(() => new MemoryQueue<Message.Message>(queueDxn), [queueDxn]);
+
+  const model = useQueueModelAdapter(renderByline([]), queue);
+  const handleSegments = useCallback<TranscriberProps['onSegments']>(
+    async (blocks) => {
+      void queue?.append([
+        Obj.make(Message.Message, {
+          created: new Date().toISOString(),
+          sender: actor,
+          blocks,
+        }),
+      ]);
+    },
+    [queue],
+  );
+
+  const transcriber = useTranscriber({
+    audioStreamTrack: track,
+    onSegments: handleSegments,
+    transcriberConfig,
+    recorderConfig,
+  });
+
+  // Normalize sentences.
+  const normalizer = useMemo(() => {
+    if (!normalizeSentences) {
+      return;
+    }
+    const executor = new FunctionExecutor(
+      new ServiceContainer().setServices({
+        // ai: {
+        //   client: new Edge AiServiceClient({
+        //     endpoint: AI_SERVICE_ENDPOINT.REMOTE,
+        //     defaultGenerationOptions: {
+        //       model: '@anthropic/claude-3-5-sonnet-20241022',
+        //     },
+        //   }),
+        // },
+      }),
+    );
+
+    return new MessageNormalizer({
+      functionExecutor: executor,
+      queue,
+      startingCursor: { actorId: getActorId(actor), timestamp: new Date().toISOString() },
+    });
+  }, [normalizeSentences, queue, actor]);
+
+  useEffect(() => {
+    if (!normalizer) {
+      return;
+    }
+    const ctx = new Context();
+    scheduleTask(ctx, async () => {
+      await normalizer.open();
+      ctx.onDispose(async () => {
+        await normalizer.close();
+      });
+    });
+    return () => {
+      void ctx.dispose();
+    };
+  }, [normalizer]);
+
+  const manageChunkRecording = () => {
+    if (running && isSpeaking && transcriber) {
+      log.info('starting transcription');
+      transcriber?.startChunksRecording();
+    } else if ((!isSpeaking || !running) && transcriber) {
+      log.info('stopping transcription');
+      transcriber?.stopChunksRecording();
+    }
+  };
+
+  useEffect(() => {
+    const ctx = new Context();
+    scheduleTask(ctx, async () => {
+      if (transcriber && running) {
+        await transcriber.open();
+      } else if (!running) {
+        await transcriber?.close();
+      }
+      manageChunkRecording();
+    });
+
+    return () => {
+      void ctx.dispose();
+    };
+  }, [transcriber, running]);
+
+  useEffect(() => {
+    manageChunkRecording();
+  }, [isSpeaking, stream]);
+
+  return (
+    <TranscriptionStory
+      disabled={!stream}
+      model={model}
+      running={running}
+      onRunningChange={setRunning}
+      audioRef={ref}
+      onUpload={handleUpload}
+    />
+  );
+};
+
+const meta = {
+  title: 'plugins/plugin-transcription/components/FileTranscription',
+  decorators: [
+    withLayout({ layout: 'column' }),
+    withPluginManager({
+      plugins: [
+        ...corePlugins(),
+        StorybookPlugin({}),
+        ClientPlugin({
+          types: [TestItem, Person.Person, Organization.Organization, TestSchema.DocumentType],
+          onClientInitialized: ({ client }) =>
+            Effect.gen(function* () {
+              const { personalSpace } = yield* initializeIdentity(client);
+              yield* Effect.promise(() => seedTestData(personalSpace));
+            }),
+        }),
+
+        PreviewPlugin(),
+        TranscriptionPlugin(),
+      ],
+      fireEvents: [AppActivationEvents.SetupAppGraph],
+    }),
+  ],
+} satisfies Meta;
+
+export default meta;
+
+type Story = StoryObj<typeof meta>;
+
+const TRANSCRIBER_CONFIG = {
+  transcribeAfterChunksAmount: 100,
+  prefixBufferChunksAmount: 50,
+  normalizeSentences: true,
+};
+
+const RECORDER_CONFIG = {
+  interval: 200,
+};
+
+export const Default: StoryObj<typeof AudioFile> = {
+  render: AudioFile,
+  args: {
+    detectSpeaking: true,
+    // https://learnenglish.britishcouncil.org/general-english/audio-zone/living-london
+    audioUrl: 'https://dxos.network/audio-london.m4a',
+    // textUrl: 'https://dxos.network/audio-london.txt',
+    transcriberConfig: TRANSCRIBER_CONFIG,
+    recorderConfig: RECORDER_CONFIG,
+  },
+};
+
+// TODO(mykola): Fix sentence normalization.
+// export const WithSentenceNormalization: StoryObj<typeof AudioFile> = {
+//   render: AudioFile,
+//   args: {
+//     detectSpeaking: true,
+//     normalizeSentences: true,
+//     // https://learnenglish.britishcouncil.org/general-english/audio-zone/living-london
+//     audioUrl: 'https://dxos.network/audio-london.m4a',
+//     // textUrl: 'https://dxos.network/audio-london.txt',
+//     transcriberConfig: TRANSCRIBER_CONFIG,
+//     recorderConfig: RECORDER_CONFIG,
+//   },
+// };

package/src/{components/stories → stories}/KeyWordRecognition.stories.tsx

@@ -2,22 +2,24 @@
 // Copyright 2025 DXOS.org
 //
 
-import '@dxos-theme';
-
-import { type StoryObj, type Meta } from '@storybook/react';
+import { type Meta, type StoryObj } from '@storybook/react-vite';
 import React, { useEffect, useMemo, useState } from 'react';
 
 import { log } from '@dxos/log';
-import { Toolbar } from '@dxos/react-ui';
-import { mx } from '@dxos/react-ui-theme';
-import { withLayout, withTheme } from '@dxos/storybook-utils';
+import { Panel, Toolbar } from '@dxos/react-ui';
+import { withLayout, withTheme } from '@dxos/react-ui/testing';
+import { mx } from '@dxos/ui-theme';
 
 type Word = {
   text: string;
   matched: boolean;
 };
 
-const SpeechRecognition = ({ keywords }: { keywords: string[] }) => {
+type DefaultStoryProps = {
+  keywords: string[];
+};
+
+const DefaultStory = ({ keywords }: DefaultStoryProps) => {
   const [running, setRunning] = useState(false);
   const [matchingWords, setMatchingWords] = useState<Word[]>(keywords.map((word) => ({ text: word, matched: false })));
 
@@ -84,45 +86,49 @@ const SpeechRecognition = ({ keywords }: { keywords: string[] }) => {
   }, [running, recognition]);
 
   return (
-    <div className={mx('flex flex-col gap-2 w-[30rem]')}>
-      <Toolbar.Root>
-        <Toolbar.IconButton
-          iconOnly
-          icon={running ? 'ph--pause--regular' : 'ph--play--regular'}
-          label={running ? 'Pause' : 'Play'}
-          onClick={() => setRunning((running) => !running)}
-        />
-      </Toolbar.Root>
-
-      <div className='flex flex-wrap'>
-        {matchingWords.map((word, index) => (
-          <span
-            key={index}
-            className={mx(
-              'px-3 py-1 m-2 rounded border',
-              word.matched ? 'bg-red-200 text-red-800 border-red-300' : 'bg-gray-200 text-gray-600 border-gray-300',
-            )}
-          >
-            {word.text}
-          </span>
-        ))}
-      </div>
-
-      <span>{transcript || (running ? 'Listening...' : 'Click play to start speech recognition')}</span>
-    </div>
+    <Panel.Root>
+      <Panel.Toolbar asChild>
+        <Toolbar.Root>
+          <Toolbar.IconButton
+            iconOnly
+            icon={running ? 'ph--pause--regular' : 'ph--play--regular'}
+            label={running ? 'Pause' : 'Play'}
+            onClick={() => setRunning((running) => !running)}
+          />
+        </Toolbar.Root>
+      </Panel.Toolbar>
+
+      <Panel.Content>
+        <div className='p-4 grid grid-cols-[repeat(auto-fill,minmax(6rem,1fr))] gap-2'>
+          {matchingWords.map((word, index) => (
+            <span
+              key={index}
+              className={mx(
+                'grid place-items-center p-2 aspect-square rounded-sm border border-separator',
+                word.matched && 'border-red-300',
+              )}
+            >
+              {word.text}
+            </span>
+          ))}
+        </div>
+      </Panel.Content>
+    </Panel.Root>
   );
 };
 
-const meta: Meta<typeof SpeechRecognition> = {
-  title: 'plugins/plugin-transcription/KeyWordDetection',
-  decorators: [withTheme, withLayout({ fullscreen: true, classNames: 'justify-center' })],
-};
+const meta = {
+  title: 'plugins/plugin-transcription/components/KeyWordDetection',
+  render: DefaultStory,
+  decorators: [withTheme(), withLayout({ layout: 'column' })],
+} satisfies Meta<typeof DefaultStory>;
 
 export default meta;
 
-export const SpeechRecognitionAPI: StoryObj<typeof SpeechRecognition> = {
-  render: SpeechRecognition,
+type Story = StoryObj<typeof meta>;
+
+export const SpeechRecognitionAPI: Story = {
   args: {
-    keywords: ['kai', 'computer', 'hello'],
+    keywords: ['kai', 'computer', 'hello', 'dxos'],
   },
 };