Sync from Jahmori-R/PureVoiceAI-HG via hub-sync

.github/workflows/publish.yml

@@ -1,23 +0,0 @@
-name: Sync to Hugging Face hub
-on:
-    push:
-        branches: [main]
-
-    # to run this workflow manually from the Actions tab
-    workflow_dispatch:
-
-jobs:
-    sync-to-hub:
-        runs-on: ubuntu-latest
-        steps:
-            - uses: actions/checkout@v2
-              with:
-                  fetch-depth: 0
-            - name: Add remote
-              env:
-                  HF: ${{ secrets.HF }}
-              run: git remote add space https://Jahmori-R:$HF@huggingface.co/spaces/Jahmori-R/PureVoiceAISpace
-            - name: Push to hub
-              env:
-                  HF: ${{ secrets.HF }}
-              run: git push --force https://Jahmori-R:$HF@huggingface.co/spaces/Jahmori-R/PureVoiceAISpace main

.gitignore

@@ -1,28 +0,0 @@
-
-
-
-# Virtual Environments
-.env
-.venv
-env/
-venv/
-ENV/
-
-#pyenv versions control
-.python-version
-
-
-# Ignore specific output directories
-uploads
-pretrained_models
-local
-
-
-# Audio formats from processing
-*.wav
-*.flac
-*.mp3
-
-
-# Configs
-pyrightconfig.json

README.md

@@ -9,26 +9,24 @@ python_version: 3.12.12
 app_file: app.py
 pinned: false
 license: apache-2.0
-short_description: Space to demonstrate ai logging pipeline system
+short_description: AI speech logging transcription pipeline
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
-
 # PureVoiceAI
 
-An advanced, end-to-end audio processing pipeline designed to improve the quality of transcription/note-taking. PureVoiceAI integrates machine learning models to perform audio source separation, voice activity detection, and speech transcription, turning noisy, overlapping audio into clean, attributed text.
+An end-to-end audio processing pipeline designed to improve the quality of transcription by addressing the speech overlap problem. PureVoiceAI integrates machine learning models to perform audio source separation, voice activity detection, and speech transcription, turning noisy, overlapping audio into clean, attributed text.
 
 ## Key Features
 
-- **Advanced Source Separation:** Utilizes modern model **RE-SepFormer** to isolate individual voices from noisy or overlapping audio tracks.
-- **Speaker Diarization:** Accurately determines "who spoke when," ensuring that transcribed text is correctly attributed to individual speakers.
-- **Speech Transcription:** Converts the cleaned, separated audio streams into highly accurate text.
-- **Modern AI Tooling:** Built on top of robust, industry-standard audio processing toolkit [SpeechBrain](https://speechbrain.github.io/).
+- **Advanced Source Separation:** Utilizes custom trained model **RE-SepFormer** to isolate individual voices from noisy or overlapping audio tracks.
+- **Voice Activity Detection:** Accurately determines "who spoke when," ensuring that transcribed text is correctly attributed to individual speakers.
+- **Speech Transcription:** Converts the cleaned, separated audio streams into highly accurate text given voice activity timestamps.
+- **Toolkits:** Built on top of audio processing toolkit [SpeechBrain](https://speechbrain.github.io/).
 
 ## Tech Stack
 
 - **Language:** Python
-- **Package Management:** `pip` 
+- **Package Management:** pip
 - **Audio Frameworks:** SpeechBrain
 - **Core ML Models:** RE-SepFormer, pyAnnote, OpenAI Whisper
 

app.py

@@ -14,20 +14,22 @@ from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
 from speechbrain.inference.separation import SepformerSeparation
 from pydub import AudioSegment
 
-# Disable TF32 for maximum precision (fixes the Pyannote warning)
+# Disables TF32 for to prevent lower accuracy in pyannote
 torch.backends.cuda.matmul.allow_tf32 = False
 torch.backends.cudnn.allow_tf32 = False
 
-device = "cuda:0" if torch.cuda.is_available() else "cpu"  # GPU availability check
+# Temporary Directory Intializing
 
-
-token = os.environ.get("HF_TOKEN")
 temp_dir = tempfile.gettempdir()
-
+print(f"temp directory created: {temp_dir}")
 
 @spaces.GPU
 def process_audio(mixture):
-    print(f"temp directory created: {temp_dir}")
+    # GPU processing availability check
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+
+    # Hugging Face authentication token grab
+    token = os.environ.get("HF_TOKEN")
 
     # Name of separated file folder
     sources = Path(temp_dir) / "sources"
@@ -36,10 +38,12 @@ def process_audio(mixture):
     Path.mkdir(sources, exist_ok=True)
 
     # Downloads RE-SepFormer model from huggingface buckets to space local directory
-    sync_bucket("hf://buckets/Jahmori-R/PureVoiceAISpace-storage", "./local")
+    # sync_bucket("hf://buckets/Jahmori-R/PureVoiceAISpace-storage", "./local")
+
+
 
     # ---------------------------------------------------------
-    # INITIALIZE MODELS
+    # 1. INITIALIZE MODELS
     # ---------------------------------------------------------
 
     print("Intializing models...")
@@ -47,28 +51,22 @@ def process_audio(mixture):
     # Loading Resepformer Model
     try:
         separator = SepformerSeparation.from_hparams(
-            source=f"./local/resepformer-librimix2spk",
+            source=f"Jahmori-R/resepformer-librimix2spk",
             hparams_file="hyperparams.yaml",
-            run_opts={"device": "cuda"},
+            run_opts={"device": device},
         )
-
-        # separator = SepformerSeparation.from_hparams(
-        #     source="Jahmori-R/resepformer-librimix2spk",
-        #     savedir=f"./pretrained_models/resepformer-librimix2spk",
-        #     # run_opts={"device": "cuda"},
-        #     # fetch_config=FetchConfig(token=True)
-        # )
-
         print("RE-SepFormer model successfully loaded")
     except Exception as e:
         print(f"Failed to load RE-SepFormer model. Error: {e}")
         raise SystemExit()
 
-    # #Loading Pyannote Model
+    # Loading Pyannote Model
+
     try:
         diarize = Pipeline.from_pretrained(
             "pyannote/speaker-diarization-community-1", token=token
         )
+        # Assigning GPU processing to model
         diarize.to(torch.device("cuda"))
         print("Pyannote model successfully loaded")
     except Exception as e:
@@ -76,8 +74,7 @@ def process_audio(mixture):
         raise SystemExit()
 
     # Loading Whisper Model
-    # Prevents WhisperAI logging in cmd
-    transformers.logging.set_verbosity_error()
+    transformers.logging.set_verbosity_error() # Prevents WhisperAI logging in cmd line
     try:
         transcription_model_id = "openai/whisper-large-v3-turbo"
         torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
@@ -108,15 +105,16 @@ def process_audio(mixture):
     # LOAD & SEPARATE AUDIO
     # ---------------------------------------------------------
 
-    # Importing mixed audio source
 
-    # Calling ReSepformer model onto audio source to separate
+    # Calling ReSepformer model onto audio source
     assert separator is not None, "Failed to load the Sepformer model."
 
+    # Running mixed audio onto separation model
     print("Separating Audio Mixture")
     est_sources = separator.separate_file(mixture)
     print("Mixture separated")
 
+    # Saving separated sources to directory
     torchaudio.save(
         f"{sources}/speaker_1.wav", est_sources[:, :, 0].detach().cpu(), 16000
     )
@@ -125,17 +123,17 @@ def process_audio(mixture):
     )
 
     # ---------------------------------------------------------
-    # DIARIZATION ON FILES
+    # VAD + TRANSCRIPTION ON SOURCES
     # ---------------------------------------------------------
 
     # Dictionary to store speaker information
     unfiltered_log = [[], []]
-    overlap_durations = []
 
     # Storing both speakers
     file_paths = sorted(
+      # Glob finds all similar files
         glob.glob(f"{sources}/speaker_*.wav")
-    )  # Glob finds all similar files
+    )
 
     spk_1 = file_paths[0]
     spk_2 = file_paths[1]
@@ -148,7 +146,7 @@ def process_audio(mixture):
         print(f"Diarizing {speaker_key}")
         diarization = diarize(path, num_speakers=1)
 
-        # Storing speaker voice activity timestamp to dictionary
+        # Storing speaker voice activity timestamps to dictionary
         for turn, _ in diarization.speaker_diarization:
             unfiltered_log[i].append(
                 {
@@ -158,6 +156,8 @@ def process_audio(mixture):
                 }
             )
 
+
+        # Creating audio segment from detected voice activity
         audio_seg = AudioSegment.from_file(path)
 
         print("Transcribing Chunks...")
@@ -168,10 +168,11 @@ def process_audio(mixture):
             end_ms = segment["end"] * 1000
             chunk = audio_seg[start_ms:end_ms]
 
-            # Converting pydub AudioSegment to numpy array
+            # Converting pydub AudioSegment to numpy array for whisper model
             samples = np.array(chunk.get_array_of_samples())
             samples = samples.astype(np.float32) / 32768.0
 
+            # Generating transcription given chunk segment
             transcription = transcribe(
                 samples,
                 generate_kwargs={
@@ -179,6 +180,7 @@ def process_audio(mixture):
                     "condition_on_prev_tokens": True,
                 },
             )
+            # Appending each segment transcription to dictionary
             segment["transcription"] = transcription
 
     # Combined both speaker logs
@@ -187,26 +189,11 @@ def process_audio(mixture):
     # Sorting by start time of timestamps
     filtered_log = sorted(combined_log, key=lambda x: x["start"])
 
-    # Calculating the overlap between each segment to assign appropriate speaker to segment
-    # for log, log_next in zip(filtered_log, filtered_log[1:]):
-    #     current_start = log["start"]
-    #     current_end = log["end"]
-
-    #     next_start = log_next["start"]
-    #     next_end = log_next["end"]
-
-    #     overlap_start = max(current_start, next_start)
-    #     overlap_end = min(current_end, next_end)
-    #     overlap = max(0.0, overlap_end - overlap_start)
-    #     overlap_durations.append(overlap)
-
-    # Overlkap formula = Min(end 1, end 2) - Max(start 1, start 2)
-
     print("Task Finished")
     return (spk_1, spk_2, filtered_log)
 
 
-# css for interface modification
+# Custom css for gradio interface
 css = """
 
 #centered-row {
@@ -215,6 +202,7 @@ css = """
 }
 """
 
+# Gradio layout configuration
 with gr.Blocks(css=css) as demo:
     gr.Markdown("PureVoiceAI")
     with gr.Row(variant="panel"):
@@ -242,4 +230,4 @@ with gr.Blocks(css=css) as demo:
     )
 
 if __name__ == "__main__":
-    demo.launch(allowed_paths=[temp_dir], theme=gr.themes.Soft())
+        demo.launch(allowed_paths=[temp_dir], theme=gr.themes.Soft())

pretrained_models/resepformer-librimix2spk/decoder.ckpt

@@ -1 +0,0 @@
-/home/jahmori/.cache/huggingface/hub/models--Jahmori-R--resepformer-librimix2spk/snapshots/b1986fdc97f0dcfd2f738fb8083c7acb661ab783/decoder.ckpt

pretrained_models/resepformer-librimix2spk/encoder.ckpt

@@ -1 +0,0 @@
-/home/jahmori/.cache/huggingface/hub/models--Jahmori-R--resepformer-librimix2spk/snapshots/b1986fdc97f0dcfd2f738fb8083c7acb661ab783/encoder.ckpt

pretrained_models/resepformer-librimix2spk/hyperparams.yaml

@@ -1 +0,0 @@
-/home/jahmori/.cache/huggingface/hub/models--Jahmori-R--resepformer-librimix2spk/snapshots/b1986fdc97f0dcfd2f738fb8083c7acb661ab783/hyperparams.yaml

pretrained_models/resepformer-librimix2spk/masknet.ckpt

@@ -1 +0,0 @@
-/home/jahmori/.cache/huggingface/hub/models--Jahmori-R--resepformer-librimix2spk/snapshots/b1986fdc97f0dcfd2f738fb8083c7acb661ab783/masknet.ckpt

requirements.txt

@@ -37,8 +37,8 @@ frozenlist==1.8.0
 fsspec==2026.7.0
 future==1.0.0
 googleapis-common-protos==1.75.3
-gradio==6.28.0
-gradio_client==2.7.1
+gradio==6.26.0
+gradio_client==2.6.1
 greenlet==3.5.6
 groovy==0.1.2
 grpcio==1.84.0