STT4SG Replicate

This repo is now a Cog/Replicate wrapper around i4Ds/stt4sg-transcribe, with the pipeline vendored locally and a repo-local model/cache layout so the same assets are available both on the host and inside the built image.

What Changed

• Replaced the old WhisperX-only predictor with the STT4SG transcription pipeline.
• Added a uv project (pyproject.toml, uv.lock once generated) for local development.
• Added repo-local cache helpers so Whisper, alignment, Silero, and pyannote assets can live under model-store/.
• Exposed the upstream pipeline parameters through Cog inputs instead of hardcoding a narrow set.

Local Layout

Downloaded assets are stored locally in this repository:

model-store/
├── whisper/        # local faster-whisper model directories
├── alignment/      # torchaudio / HF alignment models
├── huggingface/    # pyannote and other HF caches
├── torch/          # torch / torchaudio cache
├── xdg/            # xdg cache used by some backends
├── speechbrain/    # optional speechbrain cached models
└── nemo/           # optional nemo assets

That directory is copied into the Cog build context, so pre-downloaded assets are available inside the container without extra runtime downloads.

Setup

Create the local environment:

uv venv
source .venv/bin/activate
uv sync

Prefetch the default local assets:

uv run python scripts/download_assets.py

Notes:

• Default Whisper model: i4ds/daily-brook-134
• Default alignment model: VOXPOPULI_ASR_BASE_10K_DE
• Silero is prefetched automatically.
• Pyannote downloads require a Hugging Face token or an existing local HF login. The script will use your saved local login if one exists.

You can prefetch additional Whisper models too:

uv run python scripts/download_assets.py \
  --whisper-model i4ds/daily-brook-134 \
  --whisper-model Systran/faster-whisper-large-v3

Local CLI Usage

Run the vendored pipeline directly:

uv run main.py 97_Brugg.mp3

Example with explicit parameters:

uv run main.py 97_Brugg.mp3 \
  --model i4ds/daily-brook-134 \
  --vad-method silero \
  --batch-size 8 \
  --no-alignment

Cog Usage

Build the image:

cog build

Run the included test file:

cog predict \
  -i audio_file=@97_Brugg.mp3 \
  -i model=i4ds/daily-brook-134 \
  -i use_vad=true \
  -i vad_method=silero \
  -i use_alignment=true

The predictor returns:

• srt_file
• srt_content
• detected_language
• duration_seconds
• num_segments
• logs_zip when log saving is enabled

Cog Inputs

The Cog predictor surfaces the upstream runtime controls:

• audio_file
• model
• language
• task
• log_progress
• use_vad
• vad_method
• vad_params
• diarization
• diarization_method
• diarization_params
• num_speakers
• min_speakers
• max_speakers
• use_alignment
• alignment_model
• batch_size
• device
• compute_type
• include_speaker_labels
• save_logs
• hf_token

vad_params and diarization_params are JSON strings so backend-specific parameters can be passed through unchanged.

Replicate Push

cog login
cog push r8.im/your-username/stt4sg-replicate

Verification Target

The intended smoke test for this repo is:

cog predict -i audio_file=@97_Brugg.mp3

If pyannote assets are not available locally, keep diarization=false and use vad_method=silero for the basic transcription path.