Class WhisperDecodingTask

constructor

• new WhisperDecodingTask(
      model: WhisperGraph,
      tokenizer: WhisperTokenizer,
      options: WhisperDecodingOptions,
  ): WhisperDecodingTask

  Parameters

  • model: WhisperGraph
  • tokenizer: WhisperTokenizer
  • options: WhisperDecodingOptions

  Returns WhisperDecodingTask

ReadonlyinitialTokens

Readonlymodel

ReadonlynCtx

ReadonlynGroup

Readonlyoptions

ReadonlysampleBegin

ReadonlysampleLen

ReadonlysotIndex

ReadonlysotSequence

Readonlytokenizer

ReadonlytokenizerLanguage

promptBudget

• get promptBudget(): number

  Prompt-token budget the transcribe loop may carry (n_ctx/2 − 1).

  Returns number

useFast

• get useFast(): boolean

  Whether this task decodes on the optimized path.

  Returns boolean

applyFilters

• applyFilters(logits: MlxArray, tokens: number[][]): MlxArray

  Apply the configured filters to logits [B, V] f32 → new array.

  Parameters

  • logits: MlxArray
  • tokens: number[][]

  Returns MlxArray

detectLanguage

• detectLanguage(
      audioFeatures: MlxArray,
  ): { code: string; probs: Record<string, number>; token: number }

  detect_language on encoded audio features [1, 1500, D].

  Parameters

  • audioFeatures: MlxArray

  Returns { code: string; probs: Record<string, number>; token: number }

encode

• encode(mel: MlxArray): MlxArray

  Encoder for this task's path.

  Parameters

  • mel: MlxArray

  Returns MlxArray

run

• run(melOrFeatures: MlxArray): WhisperDecodingResult

  Run on a mel window [1, 3000, n_mels] (already the model dtype) or pre-encoded audio features [1, 1500, D].

  Parameters

  • melOrFeatures: MlxArray

  Returns WhisperDecodingResult