outline set up

pyproject.toml

@@ -17,6 +17,11 @@ classifiers = [
     "Operating System :: OS Independent",
 ]
 
+dependencies = [
+    "git+https://github.com/OpenPecha/fast-antx.git",
+]
+
+
 [project.optional-dependencies]
 dev = [
     "pre-commit",

src/stt_data_with_llm/LLM_post_corrector.py

@@ -0,0 +1,2 @@
+def get_LLM_corrected_text(inference_text, reference_text):
+    pass

src/stt_data_with_llm/audio_parser.py

@@ -0,0 +1,5 @@
+def get_audio(audio_url):
+    pass
+
+def get_split_audio(audio_data, AUDIO_SEG_LOWER_LIMIT, AUDIO_SEG_UPPER_LIMIT, full_audio_id):
+    pass

src/stt_data_with_llm/catalog_parser.py

@@ -0,0 +1,2 @@
+def parse_catalog(catalog):
+    pass

src/stt_data_with_llm/config.py

@@ -0,0 +1,2 @@
+AUDIO_SEG_UPPER_LIMIT = 8
+AUDIO_SEG_LOWER_LIMIT = 2

src/stt_data_with_llm/inference_transcript.py

@@ -0,0 +1,2 @@
+def get_audio_inference_text(audio_seg_data):
+    pass

src/stt_data_with_llm/main.py

@@ -0,0 +1,68 @@
+import logging
+
+from fast_antx.core import transfer
+
+from stt_data_with_llm.catalog_parser import parse_catalog
+from stt_data_with_llm.config import AUDIO_SEG_UPPER_LIMIT, AUDIO_SEG_LOWER_LIMIT
+from stt_data_with_llm.audio_parser import get_audio, get_split_audio
+from stt_data_with_llm.LLM_post_corrector import get_LLM_corrected_text
+from stt_data_with_llm.inference_transcript import get_audio_inference_text
+
+
+logging.basicConfig(filename='./pipeline.log', level=logging.INFO)
+
+def transfer_segmentation(inference_transcript, reference_transcript):
+    reference_transcript = reference_transcript.replace("\n", " ")
+    patterns = [['segmentation', '(\n)']]
+    reference_transcript_with_inference_segmentation = transfer(inference_transcript, patterns, reference_transcript)
+    return reference_transcript_with_inference_segmentation
+
+def is_valid_transcript(inference_transcript, reference_transcript):
+    pass
+
+def post_process_audio_transcript_pairs(audio_data_info):
+    post_processed_audio_transcript_pairs = {}
+    inference_transcript = ""
+    audio_url = audio_data_info.get("audio_url", "")
+    full_audio_id = audio_data_info.get("full_audio_id", "")
+    reference_transcript = audio_data_info.get("reference_transcript", "")
+    if not audio_url:
+        return None,full_audio_id
+    audio_data = get_audio(audio_url)
+    split_audio_data = get_split_audio(audio_data, AUDIO_SEG_LOWER_LIMIT, AUDIO_SEG_UPPER_LIMIT, full_audio_id)
+    for audio_seg_id, audio_seg_data in split_audio_data.items():
+        audio_seg_inference_transcript = get_audio_inference_text(audio_seg_data)
+        inference_transcript += f'{audio_seg_inference_transcript}\n'
+    if not is_valid_transcript(inference_transcript, reference_transcript):
+        return None,full_audio_id
+    reference_transcript_with_inference_segmentation = transfer_segmentation(inference_transcript, reference_transcript)
+    inference_transcripts = inference_transcript.split("\n")
+    reference_transcripts = reference_transcript_with_inference_segmentation.split("\n")
+    for seg_walker,(audio_seg_id, audio_seg_data) in enumerate(split_audio_data.items()):
+        seg_inference_text = inference_transcripts[seg_walker]
+        seg_reference_text = reference_transcripts[seg_walker]
+        seg_LLM_corrected_text = get_LLM_corrected_text(seg_inference_text, seg_reference_text)
+        post_process_audio_transcript_pairs[audio_seg_id] = {
+            "audio_seg_data": audio_seg_data,
+            "inference_transcript": seg_inference_text,
+            "reference_transcript": seg_reference_text,
+            "LLM_corrected_text": seg_LLM_corrected_text
+        }
+    return post_processed_audio_transcript_pairs,full_audio_id
+
+
+def save_post_processed_audio_transcript_pairs(post_processed_audio_transcript_pairs, audio_data_info):
+    # Save post processed audio transcript pairs in csv
+    pass
+
+
+def get_audio_transcript_pairs(audio_transcription_catalog_url):
+    audio_transcription_datas = parse_catalog(audio_transcription_catalog_url)
+    for data_id, audio_data_info in audio_transcription_datas.items():
+        post_processed_audio_transcript_pairs,full_audio_id = post_process_audio_transcript_pairs(audio_data_info)
+        if post_processed_audio_transcript_pairs:
+            save_post_processed_audio_transcript_pairs(post_processed_audio_transcript_pairs, audio_data_info)
+        else:
+            logging.info(f"Audio data with ID {full_audio_id} has invalid transcript")
+
+