include word boundaries in the biasing

[afaiza]

When I am using the word biasing based on CTC, is there a way to include word boundaries in the biasing?

For example, I see that when I boost the word "roam" I get many false accepts for the word "roaming".

Is it possible to boost only " roam "? (Note the spaces in the boundaries)

@andrusenkoau

Thank you!

[AaronZ345]

Not as a hard right-word boundary in the current CTC word spotter.

The context file is tokenized with tokenizer.text_to_ids(), and the graph marks a match as soon as it reaches the last token of the bias phrase. SentencePiece's ▁ marker gives roam a left boundary, but the graph does not wait for the next ▁ token or end of utterance before accepting it. Adding spaces around roam in the context file therefore does not reliably distinguish it from roaming.

For the existing implementation, reduce false accepts by tuning on a held-out set: lower context_score and/or raise ctc_ali_token_weight. The latter is specifically used by the word-level CTC alignment filter to reject weak spotted hypotheses.

If an exact whole-word match is required, the spotter needs a small change: delay emission of an end state until the next non-blank token starts with ▁, or the utterance ends. Another option is to post-filter each spotted interval against get_ctc_word_alignment() and keep it only when the overlapping aligned word is exactly roam. There is no current config flag that adds this right-boundary check.