---
title: Store Dialogs — Diarization Experiments
emoji: "\U0001F3A7"
colorFrom: yellow
colorTo: blue
sdk: gradio
sdk_version: 6.9.0
python_version: "3.10"
app_file: app.py
pinned: false
short_description: Diarization and transcription experiments for store dialogs.
---

# Store Dialogs — Diarization Experiments

Private Gradio Space for running speaker diarization and transcription experiments
on store dialog recordings (Russian, surveillance cameras).

## Features

- **Single file processing** — select audio from dataset or upload, choose models, run pipeline
- **Batch mode** — process multiple files at once
- **Gold comparison** — compare predictions against Label Studio annotations (DER/WER/CER)
- **Benchmark** — automated comparison of multiple model×strategy combinations
- **Grid Search** — one-at-a-time parameter tuning for the best model

## Supported models

### Diarization
- pyannote/speaker-diarization-3.1
- NeMo MSDD (experimental, requires manual install)

### Transcription
- Whisper large-v3 (faster-whisper)
- Whisper large-v3-turbo (faster-whisper)
- Whisper Podlodka Turbo — Russian fine-tuned (transformers)
- Whisper large-v3 Russian — Common Voice 17 (transformers)
- Whisper TIES-merge Podlodka — Russian merge (transformers)
- GigaAM CTC (Sber, Russian)
- GigaAM RNNT (Sber, Russian)

## Dataset

Audio and annotations are stored in the private dataset repo:
`Niko-NN/gold-store-dialogs`

## Required secrets

| Secret | Description |
|--------|-------------|
| `HF_TOKEN` | Read access to the private dataset repo and gated models (pyannote) |

## Optional environment variables

| Variable | Default | Description |
|----------|---------|-------------|
| `STORE_DIALOGS_DATA_DIR` | — | Path to a local copy of the dataset (skips Hub download) |
| `STORE_DIALOGS_DATASET_REPO` | `Niko-NN/gold-store-dialogs` | Dataset repo ID |
| `STORE_DIALOGS_BENCHMARK_VERSION` | `v1` | Benchmark version folder |