lattice-asr 0.1.0

Hardware-adaptive multilingual ASR library for the Lattice family. Pluggable engines (Parakeet, Whisper, Distil-Whisper), optional diarization, streaming.

lattice-asr

Hardware-adaptive multilingual ASR library for the Lattice family.

Status: v0.1 implementation in progress. W1 (foundation), W2 (Whisper + LID + Transcriber), W3.1 (ParakeetMlxEngine), W3.2 (ParakeetTdtEngine), W4 (RemoteEngine + lattice-asr-server), W5 (diarization), and W6.1 (perf-gate skeleton) are landed; W3-future (WhisperCppEngine for Apple Silicon multilingual) is deferred past v0.1; W6.2 (CI workflow), W6.3 (this rewrite), and v0.1.0 (tag + PyPI) are pending. All three v0.1 perf gates clear via the wrappers on canonical hosts: C1 RTF 45.43× on Switch (Apple M4 Pro, parakeet-mlx), C2 RTF 115.12× on Cypher (RTX 2070 Turing, parakeet-tdt), C3 RTF 3.35× on Switch (faster-whisper distil-large-v3 int8). 81 r_tier tests passing. See CHANGELOG.md for the unreleased detail and docs/performance-baseline.md for the canonical baselines.

What it is

A Lattice-layer library that lifts speech-to-text from per-consumer ad-hoc plumbing into a single pluggable surface. Consumers (lattice-meetbot for meeting transcription, lattice-dictate for push-to-talk dictation, WhatsApp voice-message decryption, future ambient capture) all transcribe through the same Transcriber interface.

What it solves

• Hardware adaptivity — picks Parakeet-TDT (NeMo) on NVIDIA, parakeet-mlx on Apple Silicon, Distil-Whisper via faster-whisper on CPU-only (arm64 or x86_64 — non-GPU fallback path). Single API across all paths.
• Multilingual — dual-engine load: English route uses Parakeet (fastest, English-only); non-English route uses Whisper-large-v3-turbo or Distil-Whisper. Routing via Silero LID on the first 1.5 s of audio.
• Optional diarization — transcribe(..., diarize=True) returns segments with speaker labels (pyannote.audio CPU/GPU, NVIDIA Sortformer GPU).
• Streaming — VAD-segment-bounded partial results for ambient/meeting use.
• Remote engine — RemoteEngine adapter forwards transcription to a network endpoint speaking the lattice-asr wire protocol; lets a CPU-only laptop offload to a GPU host.
• Telemetry-injected — every call records duration, engine, language, audio length to a consumer-supplied sink. No hidden coupling.

Non-goals (v0.1)

Speaker identification (deferred to a future lattice-voiceprint), on-device fine-tuning, sub-100 ms streaming partials, audio enhancement, translation, browser/mobile bindings.

Status

Phase	State
Spec	Locked — vault 02_Projects/Lattice/lattice-asr/Specifications/2026-04-27 lattice-asr v1 - Design Spec.md
Plan	Ratified and in execution — vault 02_Projects/Lattice/lattice-asr/Plans/2026-05-08 lattice-asr v0.1 - Implementation Plan.md
W1 Foundation	Landed (hardware probe, types, telemetry, config)
W2 Whisper + LID + Transcriber MVP	Landed
W3.1 ParakeetMlxEngine (Apple Silicon EN, MLX runtime)	Landed — verified Switch RTF 45.43× via wrapper
W3.2 ParakeetTdtEngine (NVIDIA EN, NeMo runtime)	Landed — verified Cypher RTX 2070 RTF 115.12× via wrapper
W3-future WhisperCppEngine (Apple Silicon multilingual)	Deferred past v0.1
W4 RemoteEngine + lattice-asr-server	Landed
W5 Diarization	Landed (pyannote + sortformer adapters + Transcriber wire-in; real-model exercise gated on HF_TOKEN / NeMo)
W6 Ship (perf, CI, release)	Partial — W6.1 perf-gate skeleton landed and all three C1/C2/C3 gates pass via wrappers on canonical hosts; W6.2 CI workflow + W6.3 README rewrite + v0.1.0 tag pending. Canonical hosts: C1+C3 on Switch (Apple M4 Pro Mac mini), C2 on Cypher (Linux + RTX 2070).
First consumer	lattice-dictate (planned)
Second consumer	lattice-meetbot (refactored transcription path)

License

Apache-2.0 (Lattice default). See LICENSE.

Family

Part of the Lattice family. Sibling libraries: lattice-meetbot, lattice-meeting, lattice-watch, lattice-dictate (planned), lattice-recall (planned).