add fastconformer model for decoder timestamp

Signed-off-by: biscayan <[email protected]>

nemo/collections/asr/parts/utils/decoder_timestamps_utils.py

@@ -332,7 +332,7 @@ def set_asr_model(self):
 
         To assign a proper decoding function for generating timestamp output,
         the name of .nemo file should include the architecture name such as:
-        'quartznet', 'conformer', and 'citrinet'.
+        'quartznet', 'conformer', 'fastconformer' and 'citrinet'.
 
         decoder_delay_in_sec is the amount of delay that is compensated during the word timestamp extraction.
         word_ts_anchor_offset is the reference point for a word and used for matching the word with diarization labels.
@@ -347,6 +347,17 @@ def set_asr_model(self):
             self.word_ts_anchor_offset = if_none_get_default(self.params['word_ts_anchor_offset'], 0.12)
             self.asr_batch_size = if_none_get_default(self.params['asr_batch_size'], 4)
             self.model_stride_in_secs = 0.02
+
+        elif 'fastconformer' in self.ASR_model_name.lower():
+            self.run_ASR = self.run_ASR_BPE_CTC
+            self.encdec_class = EncDecCTCModelBPE
+            self.decoder_delay_in_sec = if_none_get_default(self.params['decoder_delay_in_sec'], 0.08)
+            self.word_ts_anchor_offset = if_none_get_default(self.params['word_ts_anchor_offset'], 0.12)
+            self.asr_batch_size = if_none_get_default(self.params['asr_batch_size'], 16)
+            self.model_stride_in_secs = 0.08
+            # fastConformer requires buffered inference and the parameters for buffered processing.
+            self.chunk_len_in_sec = 5
+            self.total_buffer_in_secs = 25
 
         elif 'conformer' in self.ASR_model_name.lower():
             self.run_ASR = self.run_ASR_BPE_CTC