ToolUp.Voice.Core 0.22.0

ToolUp.Voice

Voice-to-text for ToolUp applications: a portable transcription abstraction, two server-side provider implementations behind it, and a client-side mic affordance attachable to any text input.

The family is four packages, each opt-in and composed independently:

The abstraction

type ITranscriptionProvider =
    abstract ProviderId: string
    abstract SupportsStreaming: bool
    abstract Transcribe: TranscriptionRequest -> Async<Result<Transcript, TranscriptionError>>
    abstract OpenSession:
        (StreamingHypothesis -> unit) * string option ->
            Async<Result<ITranscriptionSession, TranscriptionError>>

Failures are data (TranscriptionError), every boundary is Async, and the provider is stateless between Transcribe calls — the six portability rules (GP 12) are audited in TECHNICAL_GUIDE.md.

Transcript.plainText is the "just give me the words" projection for dropping recognised text straight into an input:

let words = Transcript.plainText transcript

Server: composing a provider (BYOK)

Both server providers read their API key from ISecretStore in the _platform scope — exactly the pattern the Claude / OpenAI / Gemini AI providers use — so a rotated key flows through without reconstruction and no credential is ever read from an env var directly.

// Azure AI Speech — key at _platform / "azure-speech-key" (+ region)
let provider = AzureSpeechTranscriptionProvider.create secretStore

// OpenAI Whisper — key at _platform / "openai-api-key"
let provider = WhisperTranscriptionProvider.create secretStore

let! result = provider.Transcribe (TranscriptionRequest.create "audio/webm" audioBytes)

Client: the mic affordance

VoiceInput.micButton attaches a mic toggle to any text input. Two capture modes:

• local — the browser Web Speech API (SpeechRecognition) when available: zero server round-trip, no key, interim results stream straight into the input.
• remote — MediaRecorder captures a clip, uploads it to a server endpoint backed by the composed ITranscriptionProvider, and commits the returned transcript.

Mode resolves to local if available, unless the config forces remote. See TECHNICAL_GUIDE.md for the decision table.

An opt-in mic can be wired into the AI chat prompt box with one line (default off, zero visual change when unset) via the ToolUp.AI.Client prompt-accessory registration seam:

ToolUp.Voice.Client.VoiceInput.registerPromptMic VoiceCaptureMode.Auto

Zero cost when unused (GP 13)

Nothing composes by default. A deployment that references none of these packages has no voice surface, no server endpoint, and no change to the AI chat prompt box.

Out of scope

Text-to-speech, speaker diarisation, wake-word detection, and stored-media batch transcription pipelines. This family provides live capture into text inputs and the ITranscriptionProvider seam; stored-media caption/transcript generation is a separate media-side concern.

Licensed Apache-2.0.