desphere

Flatten a sphere. desphere is a clean-room, MIT-licensed, zero-dependency transcoder from NIST SPHERE audio (the .sph format used by TIMIT, WSJ, Switchboard, and friends) to plain RIFF/WAV. The CLI is sph2wav.

No tooling, no upload: a browser page converts one or many .sph files to WAV or FLAC entirely client-side (WASM) — live at https://ucpresearch.github.io/desphere/ (source in web/). Multiple files come back as a .zip. Prefer sph2pipe/ffmpeg/sox if you have them; the page is for when you don't.

pip install desphere            # pure Python, zero deps, works anywhere
pip install "desphere[fast]"    # + optional Rust accelerator (big shorten files)

sph2wav utterance.sph             # -> utterance.wav
sph2wav utterance.sph out.wav
sph2wav --info utterance.sph      # inspect the SPHERE header
sph2wav utterance.sph -           # WAV to stdout

from desphere import read_sphere, transcode, transcode_bytes

# Streaming API (the reference):
header, data = read_sphere("utterance.sph")
with open("utterance.wav", "wb") as f:
    transcode(header, data, f)

# One-shot bytes API:
wav_bytes = transcode_bytes(open("utterance.sph", "rb").read())

Both APIs transparently use the Rust accelerator when installed (and fall back to pure Python) — same bytes out either way. With desphere[fast], the heavy kernels (shorten decode, G.711 expansion) run in Rust; pure-Python PCM is already C-speed. The CLI passes a stray .wav through unchanged (with a warning), so pointing it at an already-decoded file is harmless.

Why

libsndfile/soundfile cannot read SPHERE (especially shorten-compressed SPHERE), and the tools that can — sph2pipe, the original shorten, FFmpeg's decoder — are GPL/LGPL or otherwise awkwardly licensed. desphere is a permissively licensed, dependency-free reimplementation.

Clean-room policy

desphere is implemented only from:

• The public NIST SPHERE format description (ASCII header, typed object fields).
• ITU-T G.711 for μ-law / a-law.
• Tony Robinson (1994), SHORTEN: Simple lossless and near-lossless waveform compression (CUED/F-INFENG/TR.156) for the shorten algorithm.
• Black-box testing: running sox / ffmpeg / sph2pipe as binaries and comparing only their output — never reading their source.

Copyright protects source code, not file formats or algorithms (the idea/expression dichotomy), so consulting prose/tabular descriptions of a format — even third-party ones — is fine. The one rule we never break: we do not read GPL/LGPL source code, and we do not use writeups that embed verbatim GPL source. Authoritative non-source references (the TR, ITU specs, NIST docs) are the primary sources; prose writeups are at most a cross-check.

Supported matrix

The design principle is support the obvious lossless path first, and fail loudly on anything not yet validated — never emit a plausible-but-wrong WAV.

Coding	Status
pcm, 16-bit, 01/10 byte order	✅ supported
pcm, 32-bit, 01/10 byte order	✅ supported
pcm, multi-channel (interleaved)	✅ supported (validate against an oracle for exotic files)
pcm, 8-bit / 24-bit	⛔ rejected (UnsupportedFormat) — sign/packing not yet validated
ulaw / alaw (G.711, 8-bit)	✅ supported (Phase B) — verified byte-exact vs ffmpeg on all 256 codes
pcm,embedded-shorten-v2.00 (16-bit)	✅ supported (Phase C) — byte-exact vs ffmpeg, mono & stereo
ulaw,embedded-shorten (shorten type 8), incl. bitshift	✅ supported (Phase C) — byte-exact vs sph2pipe (real CALLHOME)
shorten QLPC (LPC) blocks	✅ supported (Phase C) — byte-exact vs shorten encoder + ffmpeg (orders 1–20)

Adding a coding means registering a decoder in desphere/codecs.py; until then the gate raises a precise error.

Test fixtures (the "zoo")

We don't depend on real corpus files to test: tools/make_fixtures.py generates a zoo of SPHERE variants — every byte order, bit depth, and channel count — and commits them under tests/fixtures/ with a manifest recording the exact expected output. Harder codings (μ-law/a-law, eventually shorten) are produced best-effort by driving sox/ffmpeg/shorten as black-box binaries.

python tools/make_fixtures.py     # regenerate fixtures + manifest
pytest                            # validate everything

Rust, WASM, and consumers

The Python in src/desphere is the reference spec; rust/ is a Rust port that reproduces it bit-for-bit (same fixtures), for speed and for the web. It builds three ways from one crate:

• a Rust library — a Rust client (praatfan is a likely consumer) imports it via a path/git dependency;
• WASM (wasm-pack build rust --target web --features wasm) — a WASM web app and the web/ page consume it; nothing is sent to a server;
• a Python extension desphere-native (pyo3/maturin) — the optional desphere[fast] accelerator above.

Because the consumers live in the same ucpresearch org, they depend on desphere directly (path/git); publishing to crates.io / npm is not required (and is skipped). PyPI is the one registry we target (the pure-Python desphere, plus the optional desphere-native wheels).

Development

The virtualenv lives outside the (Syncthing-synced) repo and is symlinked in:

uv venv ~/local/scr/venvs/desphere
ln -s ~/local/scr/venvs/desphere .venv
uv pip install -e ".[dev]"
pytest

# Rust port:
cd rust && cargo test            # byte-exact vs the same fixtures

Roadmap

• Phase A (done): SPHERE header + 16/32-bit PCM, lossless.
• Phase B (done): μ-law / a-law decode (ITU-T G.711).
• Phase C (done): embedded-shorten of 16-bit PCM, lossless μ-law (type 8, including the non-linear BITSHIFT remap), and QLPC (LPC) blocks — validated byte-for-byte vs ffmpeg / sph2pipe / the shorten encoder on real and synthetic streams, mono and stereo. Remaining (low priority): 8/24-bit linear PCM. See docs/STATUS.md and docs/SHORTEN.md.
• Rust port (done): a bit-for-bit Rust reimplementation in rust/ powers the desphere[fast] accelerator, the WASM web page, and any Rust client, mirroring praatfan-core-clean's Python-first-then-Rust path. The Python implementation stays the readable reference. Porting guidance: docs/RUST_PORT.md.
• Released: desphere and desphere-native are on PyPI (v0.1.0); the web page is live on GitHub Pages. Release process: RELEASING.md.