Class WhisperModel

constructor

• new WhisperModel(weights: Weights, config: ModelConfig): WhisperModel

  Parameters

  • weights: Weights
  • config: ModelConfig

  Returns WhisperModel

Readonlydims

Readonlydtype

alignmentHeads

• get alignmentHeads(): [number, number][]

  Alignment heads [(layer, head), …] from the checkpoint (word timestamps); falls back to every head of the last half of the decoder layers.

  Returns [number, number][]

fast

• get fast(): WhisperFastPath

  The optimized execution path (packages/inference/src/models/whisper/fast.ts), built lazily.

  Returns WhisperFastPath

isMultilingual

• get isMultilingual(): boolean

  Returns boolean

numLanguages

• get numLanguages(): number

  Returns number

decode

• decode(
      tokens: MlxArray,
      audioFeatures: MlxArray,
      cache: WhisperKvCache,
      opts?: { wantCrossQk?: boolean },
  ): { crossQk: (MlxArray | null)[]; logits: MlxArray }

  TextDecoder.call: tokens [B, T] int32, audioFeatures [B|1, 1500, D]. Returns logits [B, T, n_vocab] (weight dtype) and, when requested, the per-layer raw cross-attention qk for alignment. Mutates cache.

  Parameters

  • tokens: MlxArray
  • audioFeatures: MlxArray
  • cache: WhisperKvCache
  • opts: { wantCrossQk?: boolean } = {}

  Returns { crossQk: (MlxArray | null)[]; logits: MlxArray }

dispose

• dispose(): void

  Returns void

encode

• encode(mel: MlxArray, trace?: (name: string, a: MlxArray) => void): MlxArray

  AudioEncoder: mel [B, 3000, n_mels] (weight dtype) → [B, 1500, D].

  Parameters

  • mel: MlxArray
  • Optionaltrace: (name: string, a: MlxArray) => void

  Returns MlxArray

gelu

• gelu(x: MlxArray): MlxArray

  The compiled nn.gelu (shared with the faithful path).

  Parameters

  • x: MlxArray

  Returns MlxArray

layerNorm

• layerNorm(x: MlxArray, prefix: string): MlxArray

  Parameters

  • x: MlxArray
  • prefix: string

  Returns MlxArray

linear

• linear(x: MlxArray, prefix: string, bias?: boolean): MlxArray

  nn.Linear against a checkpoint weight (borrowed inputs, owned output).

  Parameters

  • x: MlxArray
  • prefix: string
  • bias: boolean = true

  Returns MlxArray

makeCache

• makeCache(): WhisperKvCache

  A fresh faithful-path decoder cache.

  Returns WhisperKvCache

weight

• weight(name: string): MlxArray

  Raw checkpoint tensor (borrowed; the fast path composes its own graphs).

  Parameters

  • name: string

  Returns MlxArray

weightT

• weightT(name: string): MlxArray

  Cached Wᵀ view (borrowed).

  Parameters

  • name: string

  Returns MlxArray