8000 GitHub - artalis-io/baffle: Bounded CPU-first native DSP for continuous full-duplex speech · GitHub
[go: up one dir, main page]

Skip to content
artalis-ioPublic

About

Bounded CPU-first native DSP for continuous full-duplex speech

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Baffle: bounded native DSP for conversational audio

CPU-first C11 audio processing. Baffle processes PCM; Hail acquires/transports audio; inference interprets/generates speech; Hull orchestrates applications. No Python, GPU, audio device API or inference framework is required by the core.

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel
ctest --test-dir build --output-on-failure
cmake --install build --prefix /your/prefix

Installed consumers use find_package(Baffle 0.1.0 EXACT CONFIG REQUIRED) and Baffle::baffle. ABI 0.1 is provisional. Static and shared builds are supported (-DBF_BUILD_SHARED=ON). Set BF_BUILD_TESTS=OFF for a C-only build; the session stress test uses C++ threads. The library creates no background threads.

Implemented: streaming windowed-sinc resampling and smooth drift adjustment, channel matrices/delays, linked AGC/limiting, classical VAD, spectral noise suppression, and optional mature SpeexDSP AEC. All processors use caller-owned storage, without heap allocation, locks, I/O or worker waits during processing.

Enable the reviewed, pinned CPU AEC explicitly:

cmake -S . -B build-aec -DCMAKE_BUILD_TYPE=Release -DBF_WITH_SPEEXDSP=ON
cmake --build build-aec --parallel
ctest --test-dir build-aec --output-on-failure

This option fetches SpeexDSP 1.2.1 at a pinned commit. Offline builds can supply BF_SPEEXDSP_SOURCE_DIR; selected files must match reviewed hashes. Default builds need no network. No GPU backend is implemented.

See the independent installed PCM consumer, optional installed-Hail consumer, API contract, architecture, backend assessment, measurements, and gate evidence. BF_WITH_HAIL=ON builds separate Baffle::hail_adapter against installed Hail. The core neither includes nor links Hail.

The unchanged original AEC quality target now passes after correcting capture DC rejection; the convergence report retains the original failure and includes separate recorded-English holdout evidence. Speech conditioning and AEC have deterministic synthetic evidence. Recorded-speech quality, physical acoustic behavior, device timing and audible interruption tails remain unqualified. Exploratory real-recording AEC trials were weak and are documented alongside synthetic results. This working DSP implementation is not a production-readiness certification. See qualification for the remaining work.

Real-world convergence adds a controlled AEC3 comparison, seeded transcript fixtures and actual external ASR/TTS/LLM contention measurements. Real-room cancellation and interruption quality remain PARTIAL; blank ASR hypotheses, epoch-reset tails and missing reviewed onset labels are retained. The optional acoustic tools build against installed public libraries without changing the core dependencies.

Gate H adds stage-by-stage WER/CER on 280 paired evaluations. Double-talk improves over raw input, while near-only regressions, model-sensitive blank hypotheses and assistant-only false transcripts remain explicit failures. Reviewed first-phone/syllable preservation remains open; no production claim.

Gate I compares eight processing variants, protected NS startup, stage ordering and assistant-only recordings. Frozen NS initialization now honors foreground protection; the default paired PCM/ASR results remain unchanged. Near-only recognition regressions, assistant-only false transcripts and weak real-room AEC remain unresolved. New CPU-only 16/48 kHz inference trials retain execution, scheduling and the 10 ms NS delay separately. Gate I remains PARTIAL.

Gate J rejects a bounded residual-coherence experiment: it fails assistant-only rejection, introduces a reference-loss speech blank, and worsens recorded false-word counts. The production DSP remains unchanged. Failed experiments and resource costs are retained; acoustic readiness remains PARTIAL.

Gate K fixes stale multichannel AEC state after reset, reports native backend recovery through existing fields, and prevents long-run signed schedule-counter overflow. Two isolated adaptation changes are rejected after path/ASR comparisons. Default adaptive parameters and buffering remain unchanged; moving-room and conversational readiness remain PARTIAL.

MIT licensed; see LICENSE and NOTICE. Hail's engineering conventions are adapted, not its audio-device implementation.

About

Bounded CPU-first native DSP for continuous full-duplex speech

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0