Merge pull request #240 from Monadical-SAS/enhancements

Enhancements

server/gpu/modal/reflector_transcriber.py

@@ -17,7 +17,7 @@ WHISPER_NUM_WORKERS: int = 1
 # Translation Model
 TRANSLATION_MODEL = "facebook/m2m100_1.2B"
 
-IMAGE_MODEL_DIR = "/root/transcription_models"
+IMAGE_MODEL_DIR = f"/root/transcription_models/{TRANSLATION_MODEL}"
 
 stub = Stub(name="reflector-transcriber")
 

server/poetry.lock

@@ -1551,6 +1551,17 @@ files = [
     {file = "ifaddr-0.2.0.tar.gz", hash = "sha256:cc0cbfcaabf765d44595825fb96a99bb12c79716b73b44330ea38ee2b0c4aed4"},
 ]
 
+[[package]]
+name = "inflection"
+version = "0.5.1"
+description = "A port of Ruby on Rails inflector to Python"
+optional = false
+python-versions = ">=3.5"
+files = [
+    {file = "inflection-0.5.1-py2.py3-none-any.whl", hash = "sha256:f38b2b640938a4f35ade69ac3d053042959b62a0f1076a5bbaa1b9526605a8a2"},
+    {file = "inflection-0.5.1.tar.gz", hash = "sha256:1a29730d366e996aaacffb2f1f1cb9593dc38e2ddd30c91250c6dde09ea9b417"},
+]
+
 [[package]]
 name = "iniconfig"
 version = "2.0.0"
@@ -2097,6 +2108,20 @@ files = [
 dev = ["pre-commit", "tox"]
 testing = ["pytest", "pytest-benchmark"]
 
+[[package]]
+name = "profanityfilter"
+version = "2.0.6"
+description = "A universal Python library for detecting and/or filtering profane words."
+optional = false
+python-versions = "*"
+files = [
+    {file = "profanityfilter-2.0.6-py2.py3-none-any.whl", hash = "sha256:1706c080c2364f5bfe217b2330dc35d90e02e4afa0a00ed52d5673c410b45b64"},
+    {file = "profanityfilter-2.0.6.tar.gz", hash = "sha256:ca701e22799526696963415fc36d5e943c168f1917e3c83881ffda6bf5240a30"},
+]
+
+[package.dependencies]
+inflection = "*"
+
 [[package]]
 name = "prometheus-client"
 version = "0.17.1"
@@ -3744,4 +3769,4 @@ multidict = ">=4.0"
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.11"
-content-hash = "b67b094055950a6b39da80dc7ca26b2a0e1c778f174016a00185d7219a3348b5"
+content-hash = "a85cb09a0e4b68b29c4272d550e618d2e24ace5f16b707f29e8ac4ce915c1fae"

server/pyproject.toml

@@ -32,6 +32,7 @@ transformers = "^4.32.1"
 prometheus-fastapi-instrumentator = "^6.1.0"
 sentencepiece = "^0.1.99"
 protobuf = "^4.24.3"
+profanityfilter = "^2.0.6"
 
 
 [tool.poetry.group.dev.dependencies]

server/reflector/llm/base.py

@@ -6,11 +6,12 @@ from typing import TypeVar
 
 import nltk
 from prometheus_client import Counter, Histogram
+from transformers import GenerationConfig
+
 from reflector.llm.llm_params import TaskParams
 from reflector.logger import logger as reflector_logger
 from reflector.settings import settings
 from reflector.utils.retry import retry
-from transformers import GenerationConfig
 
 T = TypeVar("T", bound="LLM")
 
@@ -214,6 +215,9 @@ class LLM:
             # Change ( ABC ), [ ABC ], etc. ==> (ABC), [ABC], etc.
             pattern = r"(?<=[\[\{\(])\s+|\s+(?=[\]\}\)])"
             title = re.sub(pattern, "", modified_title)
+            # Irrespective of casing changes, the starting letter
+            # of title is always upper-cased
+            title = title[0].upper() + title[1:]
         except Exception as e:
             reflector_logger.info(
                 f"Failed to ensure casing on {title=} " f"with exception : {str(e)}"
@@ -221,6 +225,29 @@ class LLM:
 
         return title
 
+    def trim_title(self, title: str) -> str:
+        """
+        List of manual trimming to the title.
+
+        Longer titles are prone to run into A prefix of phrases that don't
+        really add any descriptive information and in some cases, this
+        behaviour can be repeated for several consecutive topics. Trim the
+        titles to maintain quality of titles.
+        """
+        phrases_to_remove = ["Discussing", "Discussion on", "Discussion about"]
+        try:
+            pattern = (
+                r"\b(?:"
+                + "|".join(re.escape(phrase) for phrase in phrases_to_remove)
+                + r")\b"
+            )
+            title = re.sub(pattern, "", title, flags=re.IGNORECASE)
+        except Exception as e:
+            reflector_logger.info(
+                f"Failed to trim {title=} " f"with exception : {str(e)}"
+            )
+        return title
+
     async def _generate(
         self, prompt: str, gen_schema: dict | None, gen_cfg: dict | None, **kwargs
     ) -> str:

server/reflector/llm/llm_params.py

@@ -39,7 +39,7 @@ class FinalLongSummaryParams(LLMTaskParams):
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
         self._gen_cfg = GenerationConfig(
-            max_new_tokens=800, num_beams=3, do_sample=True, temperature=0.3
+            max_new_tokens=1000, num_beams=3, do_sample=True, temperature=0.3
         )
         self._instruct = """
         Take the key ideas and takeaways from the text and create a short
@@ -65,7 +65,7 @@ class FinalShortSummaryParams(LLMTaskParams):
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
         self._gen_cfg = GenerationConfig(
-            max_new_tokens=1300, num_beams=3, do_sample=True, temperature=0.3
+            max_new_tokens=800, num_beams=3, do_sample=True, temperature=0.3
         )
         self._instruct = """
         Take the key ideas and takeaways from the text and create a short
@@ -116,7 +116,7 @@ class TopicParams(LLMTaskParams):
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
         self._gen_cfg = GenerationConfig(
-            max_new_tokens=550, num_beams=6, do_sample=True, temperature=0.9
+            max_new_tokens=500, num_beams=6, do_sample=True, temperature=0.9
         )
         self._instruct = """
                 Create a JSON object as response.The JSON object must have 2 fields:

server/reflector/processors/audio_transcript.py

@@ -1,4 +1,6 @@
+from profanityfilter import ProfanityFilter
 from prometheus_client import Counter, Histogram
+
 from reflector.processors.base import Processor
 from reflector.processors.types import AudioFile, Transcript
 
@@ -38,6 +40,8 @@ class AudioTranscriptProcessor(Processor):
         self.m_transcript_call = self.m_transcript_call.labels(name)
         self.m_transcript_success = self.m_transcript_success.labels(name)
         self.m_transcript_failure = self.m_transcript_failure.labels(name)
+        self.profanity_filter = ProfanityFilter()
+        self.profanity_filter.set_censor("*")
         super().__init__(*args, **kwargs)
 
     async def _push(self, data: AudioFile):
@@ -56,3 +60,9 @@ class AudioTranscriptProcessor(Processor):
 
     async def _transcript(self, data: AudioFile):
         raise NotImplementedError
+
+    def filter_profanity(self, text: str) -> str:
+        """
+        Remove censored words from the transcript
+        """
+        return self.profanity_filter.censor(text)

server/reflector/processors/audio_transcript_modal.py

@@ -15,6 +15,7 @@ API will be a POST request to TRANSCRIPT_URL:
 from time import monotonic
 
 import httpx
+
 from reflector.processors.audio_transcript import AudioTranscriptProcessor
 from reflector.processors.audio_transcript_auto import AudioTranscriptAutoProcessor
 from reflector.processors.types import AudioFile, Transcript, TranslationLanguages, Word
@@ -86,7 +87,7 @@ class AudioTranscriptModalProcessor(AudioTranscriptProcessor):
             if source_language != target_language and target_language in result["text"]:
                 translation = result["text"][target_language]
             text = result["text"][source_language]
-
+            text = self.filter_profanity(text)
             transcript = Transcript(
                 text=text,
                 translation=translation,

server/reflector/processors/transcript_final_title.py

@@ -60,6 +60,8 @@ class TranscriptFinalTitleProcessor(Processor):
 
         accumulated_titles = ".".join([chunk.title for chunk in self.chunks])
         title_result = await self.get_title(accumulated_titles)
+        final_title = self.llm.trim_title(title_result["title"])
+        final_title = self.llm.ensure_casing(final_title)
 
-        final_title = FinalTitle(title=title_result["title"])
+        final_title = FinalTitle(title=final_title)
         await self.emit(final_title)

server/reflector/processors/transcript_topic_detector.py

@@ -54,9 +54,11 @@ class TranscriptTopicDetectorProcessor(Processor):
         text = self.transcript.text
         self.logger.info(f"Topic detector got {len(text)} length transcript")
         topic_result = await self.get_topic(text=text)
+        title = self.llm.trim_title(topic_result["title"])
+        title = self.llm.ensure_casing(title)
 
         summary = TitleSummary(
-            title=self.llm.ensure_casing(topic_result["title"]),
+            title=title,
             summary=topic_result["summary"],
             timestamp=self.transcript.timestamp,
             duration=self.transcript.duration,

server/tests/conftest.py

@@ -30,7 +30,7 @@ def dummy_processors():
         "reflector.processors.transcript_final_short_summary.TranscriptFinalShortSummaryProcessor.get_short_summary"
     ) as mock_short_summary:
         mock_topic.return_value = {"title": "LLM TITLE", "summary": "LLM SUMMARY"}
-        mock_title.return_value = {"title": "LLM FINAL TITLE"}
+        mock_title.return_value = {"title": "LLM TITLE"}
         mock_long_summary.return_value = "LLM LONG SUMMARY"
         mock_short_summary.return_value = {"short_summary": "LLM SHORT SUMMARY"}
 

server/tests/test_transcripts_rtc_ws.py

@@ -185,7 +185,7 @@ async def test_transcript_rtc_and_websocket(
 
     assert "FINAL_TITLE" in eventnames
     ev = events[eventnames.index("FINAL_TITLE")]
-    assert ev["data"]["title"] == "LLM FINAL TITLE"
+    assert ev["data"]["title"] == "LLM TITLE"
 
     # check status order
     statuses = [e["data"]["value"] for e in events if e["event"] == "STATUS"]
@@ -330,7 +330,7 @@ async def test_transcript_rtc_and_websocket_and_fr(
 
     assert "FINAL_TITLE" in eventnames
     ev = events[eventnames.index("FINAL_TITLE")]
-    assert ev["data"]["title"] == "LLM FINAL TITLE"
+    assert ev["data"]["title"] == "LLM TITLE"
 
     # check status order
     statuses = [e["data"]["value"] for e in events if e["event"] == "STATUS"]