ChopShop

ARCHIVED: ChopShop has moved

This repository is no longer maintained. All active development has moved to Taters — a renamed and expanded successor to ChopShop.

Please use Taters instead: https://github.com/ryanboyd/taters

Why the change?

• The new name reflects a broader scope (audio/video/text pipelines, feature extraction, and presets).
• Ongoing fixes, features, and documentation now land in the Taters repo.

If you're starting a new project or upgrading an existing one, head to Taters for the latest code and instructions.

---

A toolkit for turning messy A/V and text into clean, analysis-ready artifacts and features. Think of it as a small pit crew for your data: split → diarize → transcribe → gather text → extract features (dictionaries, archetypes, whisper embeddings) — with predictable filenames and folders. And, also, everything in-between.

Status: early WIP. It works, but expect rough edges and occasional breaking changes.

---

What it does (high level)

• Audio from video — pull each audio stream from a container into WAV.

• Diarize + transcribe — wrapper around Mahmoud Ashraf's whisper-diarization (CSV/SRT/TXT outputs).

• Per-speaker WAVs — cut a source WAV into one file per speaker using the transcript.

• Whisper encoder embeddings — segment-level embeddings (and general audio modes) via Faster-Whisper (CTranslate2).

• Text gatherer — stream/scale a CSV or folder of .txt into a single “analysis-ready” CSV (optionally grouped).

• Feature extraction

  • Dictionary / ContentCoder across any number of dictionaries → one wide CSV with stable column order.
  • Archetypes using archetypes (sentence-transformer) → one CSV mirroring your analysis-ready file name.

• Predictable outputs — if you don't provide an output path, ChopShop writes to ./features/<kind>/<filename>.csv, where <filename> comes from your analysis-ready CSV (so grouping/concat choices are visible in the name).

---

The API you'll use

ChopShop exposes namespaced sub-APIs for clarity:

from chopshop import ChopShop
cs = ChopShop()

# Audio
wav_paths = cs.audio.extract_wavs_from_video(input_path="input.mp4", output_dir="audio_out/")
tp = cs.diarizer.with_thirdparty(audio_path=wav_paths[0], out_dir="transcripts/", whisper_model="small", device="cuda")
cs.audio.split_wav_by_speaker(source_wav=wav_paths[0], transcript_csv=tp["csv"], out_dir="per_speaker/")

# Embeddings (transcript-driven OR general-audio)
cs.audio.export_whisper_embeddings(source_wav=wav_paths[0], transcript_csv=tp["csv"])          # segment CSV
cs.audio.export_whisper_embeddings(source_wav=wav_paths[0], strategy="nonsilent", aggregate="mean")  # general audio

# Text gather → Dictionaries
feat_csv = cs.text.analyze_with_dictionaries(
    csv_path="transcripts/session.csv",
    dict_paths=["dictionaries/LIWC-22.dicx", "dictionaries/empath-default.dicx"],
    text_cols=["text"], id_cols=["speaker"], group_by=["speaker"], delimiter=",",
)

# Text gather → Archetypes
arch_csv = cs.text.analyze_with_archetypes(
    csv_path="transcripts/session.csv",
    archetype_csvs=["dictionaries/archetypes/Suicidality.csv", "dictionaries/archetypes/Resilience.csv"],
    text_cols=["text"], id_cols=["speaker"], group_by=["speaker"], delimiter=",",
)

Default output locations If you omit out_features_csv, ChopShop writes to:

• Dictionaries → ./features/dictionary/<analysis_ready_filename>.csv
• Archetypes → ./features/archetypes/<analysis_ready_filename>.csv
• Whisper embeddings → ./features/whisper_embed/<analysis_ready_filename>.csv

The <analysis_ready_filename> comes from the text-gather step (e.g., dataset_grouped_speaker.csv), or from your provided analysis_csv.

---

CLI (quick hits)

Anything you can do in Python... well, you can also run from the terminal.

# Gather text from a CSV (auto-named output if --out omitted)
python -m chopshop.helpers.text_gather \
  --csv transcripts/session.csv \
  --text-col text --group-by speaker --delimiter , --encoding utf-8-sig

# Diarization (wrapper; writes CSV/SRT/TXT under out_dir/<basename>/)
python -m chopshop.audio.diarize_with_thirdparty \
  --audio_path audio/session_a1.wav --out_dir transcripts/ --whisper_model small --device cuda --num_speakers 2

# Whisper embeddings (general audio; nonsilent with mean pool)
python -m chopshop.audio.extract_whisper_embeddings \
  --source_wav audio/session_a1.wav \
  --strategy nonsilent --aggregate mean --output_dir features/whisper_embed/

---

Installation

A fresh virtual environment is strongly recommended.

python -m venv venv-chopshop
source venv-chopshop/bin/activate

Quick path (when available)

pip install "chopshop[diarization,cuda]"

Then install the three git extras used by the diarization wrapper:

pip install git+https://github.com/MahmoudAshraf97/demucs.git
pip install git+https://github.com/oliverguhr/deepmultilingualpunctuation.git
pip install git+https://github.com/MahmoudAshraf97/ctc-forced-aligner.git

Install PyTorch built for CUDA 12.4 (the stack ChopShop targets):

pip install --force-reinstall --no-cache-dir \
  torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 \
  --index-url https://download.pytorch.org/whl/cu124

And ensure FFmpeg is on your PATH (Ubuntu: sudo apt-get install ffmpeg, macOS: brew install ffmpeg).

Manual stack (same versions, explicit)

# Core pieces
pip install "faster-whisper>=1.1.0"
pip install "nemo-toolkit[asr]>=2.dev"
pip install git+https://github.com/MahmoudAshraf97/demucs.git
pip install git+https://github.com/oliverguhr/deepmultilingualpunctuation.git
pip install git+https://github.com/MahmoudAshraf97/ctc-forced-aligner.git

# cuDNN user-space libs (CUDA 12)
pip install -U nvidia-cudnn-cu12

# PyTorch for CUDA 12.4
pip install --force-reinstall --no-cache-dir \
  torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 \
  --index-url https://download.pytorch.org/whl/cu124

# Text features
pip install contentcoder archetyper

If you hit CUDA/cuDNN loader errors, it usually means the runtime and wheel builds don't match. Keep CUDA 12.4, cu124 wheels, and cuDNN 9 aligned.

---

Troubleshooting quickies

• Delimiter or encoding issues when gathering text Pass --delimiter and --encoding parameters explicitly for CSV inputs, just to be safe. If you run into any errors, try using --delimiter , and --encoding utf-8-sig as a starting point.

• Diarizer ignores --num_speakers Use the custom entrypoint (enabled by default) which wires num_speakers through properly... for now. If needed, pin min_num_speakers == max_num_speakers == N.

• cuDNN / CUDA symbol errors Mismatched CUDA/cuDNN vs wheel builds. Reinstall the cu124 PyTorch wheels and nvidia-cudnn-cu12.

• Embeddings subprocess fails Use device=cpu to rule out GPU issues; or set CHOPSHOP_DEBUG=1 to surface more logs.

---

Credits

• Diarization stack adapted from Mahmoud Ashraf's excellent whisper-diarization.
• Dictionaries via ContentCoder-Py; archetypes via archetypes (sentence-transformers). Well, okay, I wrote those. But I didn't know at the time that they'd be so handy. So... good job, former me.

---

License & status

MIT (see LICENSE). Active WIP; APIs and default paths may (read: will) shift as the project settles — release notes will most likely call out breaking changes.

Happy chopping.