Aava Refinement

Production-oriented, language-agnostic speech refinement for TTS systems.

From raw recordings to training-ready speech data—reproducible, resumable, and language-aware.

Aava Refinement turns source audio into segmented, filtered, transcribed, aligned, normalized, and SNAC-tokenized training records. It includes an SQLite-backed orchestrator, configurable language handling, and a local monitoring dashboard.

The pipeline itself is language-agnostic. Aava TTS v1 is its Persian reference deployment: the workflow was developed across nearly 2,500 hours of Persian source audio, producing more than half a million refined speech records.

At a glance

	Aava v1 audit
Raw source audio processed	2,360.1 h
Refined tokenized audio	~1,905.3 h
Final records	547,328
SNAC tokens	562,665,782
Active stages	11 + optional language filter

The numbers describe the Aava v1 training corpus and are included as an engineering reference.

Pipeline

1. Normalize media to 16 kHz mono WAV.
2. Isolate vocals with Demucs.
3. Detect speakers with pyannote.audio.
4. Segment speech with VAD.
5. score and filter audio quality.
6. Transcribe speech in a configured language or use automatic detection.
7. Apply language-aware text normalization.
8. Align text and audio.
9. Encode audio with SNAC.
10. Format Orpheus training records.
11. Organize release artifacts.

The optional language-filter stage keeps a configured target language and quarantines other detected languages or silent inputs. The bundled Persian normalizer powered Aava v1; other languages can use basic Unicode cleanup or a project-specific normalizer. See the pipeline guide for configuration, stage contracts, recovery, and hardware guidance.

Quick start

git clone https://github.com/KEYHAN-A/aava-refinement.git
cd aava-refinement

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

python db/init_db.py --fresh
python -m dataset.pipeline.health_check
python -m dataset.pipeline.run_pipeline --all --dry-run

Place source media under dataset/downloads/<category>/, then run:

python -m dataset.pipeline.run_pipeline --all --category podcasts

For the segment pipeline, set an ISO 639-1 language code or omit it for Whisper auto-detection:

python -m dataset.pipeline.segment_pipeline --all --language en
python -m dataset.pipeline.segment_pipeline --all

Progress is resumable:

python -m dataset.pipeline.run_pipeline --all --resume
python -m dashboard.app
# http://127.0.0.1:8765

Set AAVA_ROOT=/path/to/workspace when data and artifacts should live outside the repository.

Acquisition utilities

The dataset/pipeline/download/ directory contains RSS and yt-dlp helpers, including the source-specific recipes used during Aava v1 development.

Aava ecosystem

Artifact	Purpose	Access
Aava LoRA	Persian Orpheus 3B adapter	Public
Aava 3B	Merged, ready-to-load model	Public
KEYHAN-A/aava-persian-tts	Tokenized training corpus	Private

The dataset uses one conditioning label (aava) but contains multi-source, multi-speaker speech. It must not be described as a verified single-speaker dataset.

Repository map

dataset/pipeline/       orchestration, stages, recovery, acquisition
dashboard/              local monitoring UI
db/                     SQLite schema and initialization
assets/                 release visuals
requirements.txt        pipeline dependencies

Responsible use

Do not use speech systems for impersonation, fraud, harassment, or deceptive media. Minimize personal data and document the provenance of production data.

Code is licensed under Apache-2.0. See NOTICE for upstream attribution.

Citation

@software{aava_refinement_2026,
  title   = {Aava Refinement: Multilingual Speech Dataset Pipeline},
  author  = {{KEYHAN-A}},
  year    = {2026},
  version = {1.0.1},
  url     = {https://github.com/KEYHAN-A/aava-refinement}
}