@beda.software/ai-scribe-client

The client side of the AI scribe: framework-free TypeScript for web and React Native apps that capture a visit's speech and turn it into a transcript with speakers, using the AI scribe service. It holds no UI and no app state.

Installation

yarn add @beda.software/ai-scribe-client

Peer dependencies: @beda.software/remote-data, and optionally axios for the Service type.

Transcription

@beda.software/ai-scribe-client/transcription (also re-exported from the package root).

Area	What it does
connectRealtime	Opens an OpenAI Realtime transcription session over WebSocket with an ephemeral token, streams microphone chunks (sendChunk), emits every server event, and keeps each phrase's clip (takeClip) even when transcripts arrive out of order.
createDiarizer	Diarizes a visit's phrase clips in batches sent batchSeconds (15; 0 for one call per phrase) after the earliest began, or on flush(); each phrase's promise resolves with its own part. Calls run one at a time until all maxVoiceprints (1) slots hold a voiceprint, then in parallel. knownVoiceprints fill slots for the whole visit and must use the visit's speaker ids.
diarizeClip	Sends a clip and the voiceprints to recognise (KnownVoiceprint) to the AI scribe service's /v3/transcribe with diarization; llmStats lists every model call.
Transcript messages	insertMessage, getDiarizationOutcome, applyDiarization, removePhrase, splitPhraseClip and friends: functions that take a message list and return a new one, for any store.
buildWav, getLevel	A WAV from a clip's 24 kHz PCM16 samples, and a chunk's level.
parseWav	The samples of a WAV in the format buildWav writes; throws on any other file.
cutSeconds	A view of a time range of 24 kHz samples.
resampleTo24k	Linear-interpolation resample of one Float32 microphone chunk to 24 kHz.
convertFloatToInt16	Float32 samples as PCM16.
checkVoiceprint	Checks a clip of a script read aloud and cuts the voiceprint from it, through any FetchDiarization. MIN_VOICEPRINT_CLIP_SECONDS and MAX_VOICEPRINT_CLIP_SECONDS bound the clip.

The app supplies what differs per platform and backend: the microphone, the ephemeral token request, an HTTP client bound to the AI scribe service's base URL (service, any (axiosConfig) => Promise<RemoteDataResult>), how a WAV is attached to a form on React Native (AttachWav) and where messages are stored.

Vocabulary

• samples: raw audio data, 24 kHz PCM16 (Int16Array); the name of every field that holds audio;
• chunk: a piece of microphone audio streamed to the realtime server as it is recorded;
• clip: a stretch of samples cut out for one purpose: a phrase's audio, what one diarization call sends, or a script read aloud for a voiceprint;
• wav: a clip encoded as a WAV file, as it is uploaded;
• phrase: what the realtime server's voice detection cuts between pauses, keyed by its item_id;
• segment: one speaker's part of a phrase, as the diarizer returns it;
• message: one transcript entry. A phrase is one message until it is diarized, then one message per segment;
• voiceprint: a few seconds of one speaker's voice, sent with every diarization call so the diarizer recognises them across calls; checked from a clip read aloud (checkVoiceprint) or cut from the visit;
• speaker id: a letter for the whole visit, A, B, …: one per voiceprint, and one shared by every other voice. A voice that can't be told apart yet gets none, and its segment becomes a raw message;
• batch: phrase clips joined, with 1.2 s of silence between them, into one diarization call.

Commands

@beda.software/ai-scribe-client/commands (also re-exported from the package root).

Area	What it does
detectCommand	Sends a text and the commands on offer to the AI scribe service's /v3/detect-command; returns the chosen command's payload merged with the said parameters, or null.

A command's payload is echoed back and never sent to the model, so it carries the action to run, which every payload needs; a parameter may not reuse a payload key. The app decides which commands to offer and runs the one returned. DetectedCommandOf types the result per action from the commands' own parameters, as long as the list is kept as const or written inline: typing it Command<...>[] widens the parameters away.

Releasing

Publishing a GitHub release runs .github/workflows/publish.yml, which publishes to npm with provenance through npm trusted publishing (no token). A release marked as a prerelease is published under the alpha tag, any other under latest.

Development

Prerequisites: Node ^22.18.0 || ^24.11.0 || >=26 and Corepack (pins Yarn 4 via packageManager).

corepack enable
yarn install
yarn typecheck
yarn check
yarn test