Scanovich.ai-audio-call

Local-first call analytics platform with shared pipeline, pilot-ready API/UI, saved-result history, and on-prem deployment paths

Comment

Call Analytics Platform

License: MIT Python 3.12 Platform: Linux CI

Canonical public entrypoint for the repository. For the extended guide and command reference, see README_EN.md.

At a glance

flowchart LR
  voip[VoIP_or_files] --> input[input_folder]
  input --> daemon[daemon_or_web]
  daemon --> pipeline[shared_pipeline]
  pipeline --> asr[Whisper_ASR]
  asr --> llm[OpenAI_compatible_LLM]
  llm --> qa[Quality_scores]
  qa --> artifacts[output_metadata_quality]
  artifacts --> ui[browser_UI_and_API]
  • The shared single-file flow lives in src/pipeline_service.py.
  • The web/API layer lives in src/web/app.py.
  • The browser UI is a static app under src/web/static/.
  • main.py web is the primary entrypoint for the demo and pilot-ready web layer.
  • Each successful web analysis persists artifacts to output/, metadata/, and quality_analysis/, which now also power the recent-analyses view.
  • Remote OpenAI-compatible LLM endpoints are already supported by config; see docs/REMOTE_ASR_AND_LLM.md.

Current product surface

The repository currently has two practical entry modes:

  1. uv run python main.py run for the long-running daemon pipeline over input/.
  2. uv run python main.py web for a browser UI plus HTTP API over the same shared pipeline.

The web layer supports:

  • GET /healthz
  • POST /analyze
  • GET /analyses
  • GET /analyses/{result_id}
  • / for the browser UI

Quick start

If you are evaluating fit first, start with docs/EVALUATION_GUIDE.md.

1. Install

git clone https://github.com/FUYOH666/Scanovich.ai-audio-call.git call-analytics
cd call-analytics
uv sync
cp config.example.yaml config.yaml
cp branches.example.yaml branches.yaml

2. Configure the local worker

In config.yaml:

asr:
  model_preset: "auto"
  device: "cuda"

If you want to use a remote OpenAI-compatible LLM instead of a local server, point vllm.base_url and quality_analysis.base_url at that endpoint. Use placeholders in repo-tracked config and keep real hostnames in local config or environment overrides only.

3. Run the web layer

uv run python main.py web

Open http://127.0.0.1:8080.

What you can do in the UI today:

  • upload one audio file,
  • inspect transcript, classification, ASR metrics, and quality output,
  • review recent analyses that were already persisted to disk,
  • reopen an earlier analysis without re-uploading the file.

4. Protect a pilot deployment

export WEB__REQUIRE_API_KEY=true
export WEB__API_KEY=replace-with-a-strong-key
uv run python main.py web --host 0.0.0.0 --port 8080

The browser UI already supports X-API-Key through the optional API key field.

5. Run the daemon pipeline

uv run python main.py run

Use this mode when VoIP downloaders or other systems write recordings into input/.

Artifact model

The web layer and the daemon share the same persisted artifact story:

  • output/<result_id>.txt for the cleaned transcript
  • metadata/<result_id>.json for classification and ASR metrics
  • quality_analysis/individual/<result_id>.json for quality results when enabled

That persisted state is now reused directly by the recent-analyses API and UI rather than copied into a separate database.

Documentation map

Community

Local-first direction

The local-first story is already partly implemented:

  • ASR is in-process Faster-Whisper today.
  • LLM post-processing and quality analysis can already use a local or remote OpenAI-compatible server.
  • Optional HTTP ASR is a logical next adapter-based step, but it is not implemented in this pass.

Commercial support

This repository is MIT-licensed and self-hostable. If you need pilot setup, on-prem deployment, or business-specific criteria tuning, see FUNDING.md.

Built With

Share this project:

Updates