Struct HybridDecoder 

pub struct HybridDecoder { /* private fields */ }

Multi-hypothesis decoding strategy that tries multiple approaches

Implementations

impl HybridDecoder

pub fn new(config: DecodingConfig) -> Self

Create a new hybrid decoder

pub fn decode( &self, token_probs: &[Vec<f32>], initial_token: u32, vocab_size: usize, eos_token: u32, pad_token: u32, ) -> Result<Vec<u32>>

Decode with beam-search / greedy fallback (no quality gating).

pub fn decode_with_fallback( &self, token_probs: &[Vec<f32>], initial_token: u32, vocab_size: usize, eos_token: u32, no_speech_token: u32, decode_text: impl Fn(&[u32]) -> String, ) -> Result<Vec<u32>>

Decode with quality-gated temperature fallback (faster-whisper SOTA strategy).

Iterates over config.temperatures in order. At each temperature, samples tokens from the collected logits and checks:

• no_speech_prob > no_speech_threshold → return empty (silence)
• avg_log_prob < log_prob_threshold → retry at next temperature
• compression_ratio > compression_ratio_threshold → retry at next temperature

Returns the first result that passes all quality gates, or the last attempt if all temperatures are exhausted.

Arguments

• token_probs — per-step raw logits collected autoregressively from the decoder
• initial_token — first token to prepend to the output sequence
• vocab_size — vocabulary size (bounds check for no_speech_token)
• eos_token — end-of-sequence token; stops generation when sampled
• no_speech_token — token whose first-step softmax probability signals silence
• decode_text — closure that converts token IDs to a UTF-8 string for compression ratio