Fix segment and reduce block size (#53)

Author: loubbrad

amt/inference/transcribe.py

@@ -35,7 +35,7 @@
 torch._inductor.config.fx_graph_cache = True
 
 MAX_SEQ_LEN = 4096
-MAX_BLOCK_LEN = 4096
+MAX_BLOCK_LEN = 2048
 LEN_MS = 30000
 STRIDE_FACTOR = 3
 CHUNK_LEN_MS = LEN_MS // STRIDE_FACTOR
@@ -294,7 +294,7 @@ def process_segments(
     # to make sure that a sequence of the correct format is returned. Right now
     # it messes things up somehow
     if not all(_idx <= idx for _idx in eos_idxs):
-        logger.warning("Context length overflow when transcribing segment")
+        logger.warning("Context length overflow when transcribing segment(s)")
 
     results = [
         tokenizer.decode(seq[_idx, : eos_idxs[_idx] + 1])

amt/run.py

@@ -467,7 +467,7 @@ def transcribe(
 
     files_to_process = []
     for audio_path in file_paths:
-        if segments_by_audio_file.get(audio_path, None):
+        if segments_by_audio_file.get(audio_path, None) is not None:
             file_info = {
                 "path": audio_path,
                 "segments": segments_by_audio_file[audio_path],