kaldi.kws

Functions

encode_kws_disambiguation_symbols Encodes (disambiguation-symbol, utterance-id) pairs.
lattice_to_kws_index Creates an inverted KWS index of the given lattice.
optimize_kws_index Optimizes KWS index.
search_kws_index Searches given keyword (FST) inside the KWS index (FST).
kaldi.kws.encode_kws_disambiguation_symbols(index:KwsIndexVectorFst, encode_table:KwsIndexEncodeTable)

Encodes (disambiguation-symbol, utterance-id) pairs.

Encodes labels on final arcs. Replaces output labels of final arcs (utterance ids) with the encoded labels. Replaces input labels of final arcs (disambiguation symbols) with epsilons.

Parameters:
kaldi.kws.lattice_to_kws_index(clat, utterance_id, max_silence_frames=50, max_states=-1, allow_partial=True, destructive=False)[source]

Creates an inverted KWS index of the given lattice.

The output KWS index is over the KwsIndexWeight semiring (a triplet of tropical weights with lexicographic ordering). Input lattice should be topologically sorted. For details of the algorithm, see: Dogan Can and Murat Saraclar, 2011, “Lattice Indexing for Spoken Term Detection”.

Parameters:
  • clat (CompactLatticeVectorFst) – The input lattice.
  • utterance_id (int) – The integer id to use for the input lattice.
  • max_silence_frames (int) – The duration of the longest silence (epsilon) arcs allowed in the output. Longer silence arcs will be removed. If < 0, all silence arcs are kept.
  • max_states (int) – The maximum number of states allowed in the output. If <= 0, any number of states in the output is OK.
  • allow_partial (bool) – Whether to allow partial output or skip determinization if determinization fails.
  • destructive (bool) – Whether to use the destructive implementation which avoids a copy by modifying the input lattice.
Returns:

The output KWS index FST.
Return type:

KwsIndexVectorFst
kaldi.kws.optimize_kws_index(index:KwsIndexVectorFst, max_states:int=default)

Optimizes KWS index.

Optimizes KWS index by doing encoded epsilon removal, determinization and minimization.

Parameters:
  • index (KwsIndexVectorFst) – The index FST.
  • max_states (int) – The maximum number of states allowed in the output. If <= 0, any number of states in the output is OK. Default: -1.
kaldi.kws.search_kws_index(index, keyword, encode_table, n_best=-1, detailed=False)[source]

Searches given keyword (FST) inside the KWS index (FST).

Returns the n_best results found. Each result is a tuple of (utt_id, time_beg, time_end, score).

Since keyword can be an FST, there can be multiple matching paths in the keyword and the index in a given time period. If detailed == True, stats output provides the results for all matching paths together with appropriate scores while ilabels output provides the input labels on those paths.

Parameters:
  • index (KwsIndexVectorFst) – The index FST.
  • keyword (StdVectorFst) – The keyword FST.
  • encode_table (KwsIndexEncodeTable) – The table to use for decoding output labels into utterance ids. This table is produced by encode_kws_disambiguation_symbols().
  • n_best (int) – The number of best results to return. If <= 0, all results found in the index are returned.
  • detailed (bool) – Whether to return detailed results representing individual index matches and the input label sequences for those matches. If True, output is a tuple of (results, stats, ilabels).