Provider System

The provider system is DeLive's abstraction layer for ASR backends. It unifies six different speech recognition services behind a common contract.

Architecture

ProviderRegistry (singleton)
  ├── register(entry: { info, create })
  ├── getInfo(vendorId) → ASRProviderInfo
  └── createProvider(vendorId) → ASRProvider

ASRProvider (interface)
  ├── connect(config) → Promise<void>
  ├── disconnect() → Promise<void>
  ├── sendAudio(data: Blob | ArrayBuffer) → void
  └── events: onTokens, onPartial, onFinal, onError, onFinished

BaseASRProvider (abstract class)
  ├── State machine: idle → connecting → connected → disconnecting
  ├── Event emitter (Map of callbacks)
  └── Helper methods: emitToken, emitPartial, emitFinal, etc.

WindowedBatchTranscriptionProvider<TChunk> (extends BaseASRProvider)
  ├── RollingAudioBuffer (max 45s sliding window)
  ├── Scheduling: interval or debounce
  ├── TranscriptStabilizer (commits stable text prefixes)
  └── Abstract: transcribeWindow(), resolveAudioChunk()

Provider Registry

The ProviderRegistry is a singleton that maps vendor IDs to factory functions:

providerRegistry.register({
  info: new SonioxProvider().info,
  create: () => new SonioxProvider(),
})

Six providers are registered at module load via registerDefaultProviders().

ASRProviderInfo

Each provider declares its capabilities through ASRProviderInfo:

interface ASRProviderInfo {
  id: ASRVendor
  name: string
  description: string
  type: ProviderType            // 'cloud' | 'local'
  supportsStreaming: boolean
  capabilities: ASRProviderCapabilities
  requiredConfigKeys: string[]
  supportedLanguages: string[]
  website: string
  configFields: ProviderConfigField[]
}

ASRProviderCapabilities includes:

• audioInputMode: 'media-recorder' or 'pcm16'
• audioProfile: sample rate, channels, payload format, chunk timing
• transport: execution mode, capture restart strategy
• Feature flags: supportsTranslation, supportsSpeakerDiarization, prefersTokenEvents

Event System

Providers emit events through an internal Map<eventName, callback[]>:

Event	Payload	Description
onTokens	TranscriptToken[]	Token-level updates (Soniox)
onPartial	string	Partial/in-progress text
onFinal	string	Final committed text
onError	ASRError	Error with code and message
onStateChange	ASRProviderState	Provider state changed
onFinished	—	Provider session completed cleanly

INFO

When prefersTokenEvents is true (Soniox), ProviderSessionManager does not register onPartial and onFinal listeners — it derives text from onTokens instead.

Windowed Batch Transcription

Four providers (Groq, SiliconFlow, Local OpenAI, whisper.cpp) extend WindowedBatchTranscriptionProvider:

Rolling Audio Buffer

A RollingAudioBuffer<TChunk> stores timestamped audio chunks with a maximum duration of 45 seconds. When the buffer exceeds the limit, oldest chunks are trimmed.

Scheduling

• Interval mode (Groq, SiliconFlow, whisper.cpp): setInterval every 1500ms triggers retranscription
• Debounce mode (Local OpenAI): each new audio chunk resets a 1200ms setTimeout (same value as transcribeIntervalMs)

Stabilization

TranscriptStabilizer compares successive transcription results:

1. Finds the longest common prefix between the new and previous transcription
2. Commits the stable prefix as final text
3. Emits the remainder as partial text

This prevents text flickering in the UI when the same audio window produces slightly different transcriptions.

Text Overlap Handling

When the rolling buffer advances, buildWindowedTranscriptSnapshot strips up to 200 characters of overlap between the committed text tail and the new window transcription to avoid duplicate content.

Audio Pipeline

CaptureManager selects the audio path based on capabilities.audioInputMode:

MediaRecorder Path

Used by Soniox and Local OpenAI-compatible.

1. getDisplayMedia({ audio: true }) → MediaStream
2. MediaRecorder with audio/webm;codecs=opus (or fallback)
3. ondataavailable every 100ms → Blob chunks
4. Chunks forwarded to provider.sendAudio(blob)

PCM16 Path

Used by Volcengine, Groq, SiliconFlow, and whisper.cpp.

1. getDisplayMedia({ audio: true }) → MediaStream
2. AudioProcessor with AudioWorklet (preferred) or ScriptProcessorNode (fallback)
3. Resampled to 16 kHz mono
4. Float32 → Int16 conversion → ArrayBuffer chunks
5. Chunks forwarded to provider.sendAudio(buffer)