Struct DecodeContext 

pub struct DecodeContext<'a> {
    pub prompt_tokens: &'a [u32],
    pub language_token: u32,
    pub task_token: u32,
    pub sot_token: u32,
    pub eot_token: u32,
    pub no_speech_token: u32,
    pub timestamp_begin_token: u32,
    pub notimestamps_token: u32,
    pub max_new_tokens: usize,
    pub no_speech_threshold: f32,
}

Fields

prompt_tokens: &'a [u32]

Tokens from the previous committed utterance fed as a prompt prefix before <|sot|>. Empty slice for the first window.

language_token: u32

task_token: u32

Task token (<|transcribe|> or <|translate|>). Translate is X → English only.

sot_token: u32

eot_token: u32

no_speech_token: u32

timestamp_begin_token: u32

First timestamp token ID (50364 for all current Whisper models).

notimestamps_token: u32

<|notimestamps|> token ID. Pushed as the 4th init token so the decoder produces plain text (no per-token timestamps between words). Required for greedy decode to produce coherent content on this checkpoint — greedy + timestamps-on emits mostly timestamps with little content even with ApplyTimestampRules-style filtering.

max_new_tokens: usize

Hard cap on new tokens generated (not counting prompt or init tokens).

no_speech_threshold: f32

Return empty Vec when P(no_speech) exceeds this at step 0 (default 0.6).