ragleap-voice v0.3.1 release notes

Status: pre-alpha. Open, local, MIT; no paid API in the default path. No model is verified yet and none is a default (calls need allow_unverified=True). transcribe_stream is a batch adapter with no partial transcripts; synthesize_stream is chunked playback of a finished waveform. A-law helpers are not exported. The G.711 checks follow a stated convention and make no conformance claim. Nothing is measured on real phone audio or on Python 3.13. No claim is made of being better than any other product; there is no such benchmark.

[0.3.1] - 2026-11-01

Status

Still pre-alpha. Every model is unverified and none is a default. transcribe_stream is a batch adapter (it buffers the audio and runs transcribe() once at the end, so there are no partial transcripts) and synthesize_stream is chunked playback of a finished waveform; neither has been measured on real phone audio. A-law is implemented but not exported. The G.711 checks show the encoders follow the standard's tables under a stated convention (exact decision values go to the upper interval); this is not a conformance claim. Python 3.13 has not been measured for this version.

Added

ragleap_voice.audio: alaw_encode() and alaw_decode() (ITU-T G.711 Tables 1a/1b), not exported. The decoder matches Table 1 on all 256 codes; the encoder assigns an input exactly on a decision value to the upper interval, which differs from audioop.lin2alaw on the 127 negative decision values (one code each). G.711 section 3.6 leaves that choice to the equipment specification. No conformance claim. Table checks for mu-law and A-law are in benchmarks/voice/.

Changed

transcribe_stream and synthesize_stream are now exported from ragleap_voice (from ragleap_voice import transcribe_stream, synthesize_stream). They are still a batch adapter and chunked playback of a finished waveform, as documented in docs/design/streaming.md; no partial transcripts, no latency claim. Down2Stream, Up2Stream and StreamGate stay unexported.

Install: pip install ragleap-voice==0.3.1 (engines are optional extras, e.g. ragleap-voice[whisper]).
Package: https://pypi.org/project/ragleap-voice/0.3.1/