Sadda (Pali: सद्द) — sound, voice.

An open-source toolkit for phonetics and speech-science research: a Rust engine, a Python API, and a desktop app for acoustic analysis, annotation, forced alignment, and clinical-style voice metrics.

→ Main repository: sadda-speech/sadda

A single engine with three ways in:

• Python — the primary API. pip install sadda (add [ml] for VAD + embeddings). Pre-built wheels for Linux / macOS / Windows on Python 3.10–3.13.
• Desktop app — an egui GUI with waveform, spectrogram, and measure-track lanes, tier editing, live recording, reference-distribution overlays, and an embedded Python script panel. Bundles on the Releases page.
• Rust — the sadda-engine crate underneath both.

• Acoustic DSP — STFT, spectrogram, intensity, LTAS, MFCC; pitch via Boersma / YIN / pYIN / SWIPE'; LPC formants. Multiple, separately-cited methods where the field offers more than one.
• Corpus + annotation — projects and bundles over six tier types, plus a full annotation-campaign suite: rubric-as-data, computational criteria that auto-propose annotations, annotator assignment, an inter-annotator agreement engine (Cohen's κ), a QA dashboard, and a PI lab-notebook.
• Clinical voice metrics (research & education, non-diagnostic) — jitter, shimmer, HNR, CPP / CPPS, H1–H2, GNE, and the composite AVQI / ABI indices; validated against reference implementations where one exists, clean-room from the source publications where none does.
• Reference distributions — install, query, pin, and publish normative distributions and target zones.
• ML inference — bundled Silero VAD plus a wav2vec2 / Whisper-style embedding harness via load-dynamic ONNX Runtime.
• Interop — Praat TextGrid and ELAN .eaf import/export, with documented round-trip lossiness.
• Live recording, calibration & provenance, recipes — streaming capture; absolute dB-SPL calibration; every analysis records a provenance row and can emit a runnable script + a citation list for the methods it used.

Written and under review on the main repo, shipping in the next release:

• Forced alignment & ASR — phone-level forced alignment, native espeak-ng G2P, syllabification, MFA gold-standard passthrough, and a no-transcript path (Whisper ASR → transcript → align).
• Publication-quality figure export — waveform / spectrogram / measure / tier figures to SVG, PDF, or TikZ, or straight to the clipboard.
• Adaptive speaker pitch ranges — two-pass and speaker-level (pooled / empirical-Bayes) f0-range estimation.

Public and actively developed at v0.5.x, solo part-time. Stable: corpus, dsp, clinical (research-use). Provisional: live, recipe, refdist, ml. Further out: articulatory channels (EGG / EMA / ultrasound), an experiment runner, and mobile (iOS / Android via UniFFI).

Not a medical device. The clinical-style measures are for research and education; sadda makes no diagnostic or treatment claims.

• sadda — the toolkit (engine + Python + desktop app).
• SpectroGame — phonetics training activities built on sadda.

Dual-licensed under Apache-2.0 or MIT, at your option.