add ctc only model

Signed-off-by: andrusenkoau <andrusenkoau@gmail.com>

scripts/asr_context_biasing/compute_key_words_fscore.py

@@ -122,14 +122,28 @@ def main():
         "--input_manifest", type=str, required=True, help="nemo manifest with recognition results in pred_text field",
     )
     parser.add_argument(
-        "--context_biasing_file", type=str, required=True, help="file of context biasing words/phrases with their spellings"
+        "--context_biasing_file", type=str, required=True, 
+        help="""
+             file of context biasing words/phrases with their spellings \
+             (one word/phrase per line, spellings are separated from word/phrase by dash symbol):
+             WORD1_SPELLING1
+             WORD2_SPELLING1_SPELLING2
+             ...
+             nvidia_nvidia
+             gpu_gpu_g p u
+             nvlink_nvlink_nv link
+             ...
+             alternative spellings help to improve the recognition accuracy of abbriviations and complicated words,
+             which are aften recognized as separate words by ASR model (gpu -> g p u, tensorrt -> tensor rt, and so on).
+             """
     )
 
     args = parser.parse_args()
     # use list instead of dict to preserve key words order during printing word-level statistics
     key_words_list = []
     for line in open(args.context_biasing_file).readlines():
-        item = line.strip().split("-")[0].lower()
+        item = line.strip().split("_")[0].lower()
+        assert len(item) > 1, f"word/phrase {item} does not have any spelling"
         if item not in key_words_list:
             key_words_list.append(item)
     compute_fscore(args.input_manifest, key_words_list)

scripts/asr_context_biasing/context_biasing_utils.py

@@ -25,6 +25,7 @@ def merge_alignment_with_ws_hyps(
     cb_results: List[WSHyp],
     decoder_type: str = "ctc",
     intersection_threshold: float = 30.0,
+    blank_idx: int = 0,
 ) -> str:
     """
     Merge context biasing predictions with ctc/rnnt word-level alignment.
@@ -45,7 +46,7 @@ def merge_alignment_with_ws_hyps(
         alignment_tokens = []
         prev_token = None
         for idx, token in enumerate(candidate):
-            if token != asr_model.decoder.blank_idx:
+            if token != blank_idx:
                 if token == prev_token:
                     alignment_tokens[-1] = [idx, asr_model.tokenizer.ids_to_tokens([int(token)])[0]]
                 else:

scripts/asr_context_biasing/context_graph_ctc.py

@@ -28,6 +28,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+# The script is obtained and modified from Icefall repo:
+# https://github.com/k2-fsa/icefall/blob/master/icefall/context_graph.py
+
 from collections import deque
 from typing import Dict, List, Optional, Union
 

scripts/asr_context_biasing/ctc_based_word_spotter.py

@@ -152,7 +152,7 @@ def find_best_hyps(spotted_words: List[WSHyp], intersection_threshold: int = 10)
     return best_hyp_list
 
 
-def get_ctc_word_alignment(logprob: np.ndarray, asr_model, token_weight: float = 1.0) -> List[tuple]:
+def get_ctc_word_alignment(logprob: np.ndarray, asr_model, token_weight: float = 1.0, blank_idx: int = 0) -> List[tuple]:
     """ 
     Get word level alignment (with start and end frames) based on argmax ctc predictions.
     The word score is a sum of non-blank token logprobs with additional token_weight.
@@ -174,7 +174,7 @@ def get_ctc_word_alignment(logprob: np.ndarray, asr_model, token_weight: float =
     prev_idx = None
     for i, idx in enumerate(alignment_ctc):
         token_logprob = 0
-        if idx != asr_model.decoder.blank_idx:
+        if idx != blank_idx:
             token = asr_model.tokenizer.ids_to_tokens([int(idx)])[0]
             if idx == prev_idx:
                 prev_repited_token = token_alignment.pop()
@@ -268,6 +268,7 @@ def run_word_spotter(
         logprobs: np.ndarray,
         context_graph: ContextGraphCTC,
         asr_model,
+        blank_idx: int = 0,
         beam_threshold: float = 5.0,
         cb_weight: float = 3.0,
         ctc_ali_token_weight: float = 0.5,
@@ -284,6 +285,7 @@ def run_word_spotter(
     Args:
         logprobs: CTC logprobs
         context_graph: Context-Biasing graph
+        blank_idx: blank index in ASR model
         asr_model: ASR model (ctc or hybrid-transducer-ctc)
         beam_threshold: threshold for beam pruning
         cb_weight: context biasing weight
@@ -306,8 +308,6 @@ def run_word_spotter(
     blank_threshold = np.log(blank_threshold)
     non_blank_threshold = np.log(non_blank_threshold)
 
-    blank_idx = asr_model.decoder.blank_idx
-
     for frame in range(logprobs.shape[0]):
         # add an empty token (located in the graph root) at each new frame to start new word spotting
         active_tokens.append(Token(start_state, start_frame=frame))
@@ -357,7 +357,7 @@ def run_word_spotter(
     best_hyp_list = find_best_hyps(spotted_words)
 
     # filter hyps according to word-level ctc predictions to avoid a high false accept rate
-    ctc_word_alignment = get_ctc_word_alignment(logprobs, asr_model, token_weight=ctc_ali_token_weight)
+    ctc_word_alignment = get_ctc_word_alignment(logprobs, asr_model, token_weight=ctc_ali_token_weight, blank_idx=blank_idx)
     best_hyp_list = filter_wb_hyps(best_hyp_list, ctc_word_alignment)
 
     return best_hyp_list