Context biasing to improve recognition of domain-specific vocabulary (proper nouns, product names, jargon).

make build
./build/examples/example-phrase-boost model.safetensors vocab.txt audio.wav

Phrase boosting builds a token-level trie from boost phrases and biases the decoder's log-probabilities toward trie-matched tokens during greedy decode. This steers recognition without retraining.

• boost_phrases: list of words/phrases to boost
• boost_score: log-probability bias strength (default: 5.0)

Baseline: Well, I don't wish to see it anymore, observed Phoebi, turning away...
Boosted:  Well, I don't wish to see it anymore, observed Phoebe, turning away...