Skip to content
CAI
Software that uses CAICheck a score

FluidInference/FluidAudio

60.8

Adequate · 14 November 2026

123.2k

lines of production code

Swift

primary language

5

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

Features

Add NVIDIA Canary-1B-v2 ASR backend with custom vocabulary boosting

Introduces a new beta transcription backend for the NVIDIA Canary-1B-v2 model, featuring int4, fp16, and int8 precision options with CoreML/ANE acceleration. The implementation includes a 15-second sliding-window pipeline with token-level seam merging for long-form audio, and a \CanaryKeywordBooster\ that applies custom-vocabulary boosting by reusing the existing CTC keyword spotter to detect and inject terms into the autoregressive transcript.

Sources · high confidence

Added NeMo Sortformer AMI benchmark script and documentation

Introduced a new Python benchmark tool (\nemo\_ami\_benchmark.py\) and its documentation (\README.md\) in the \Scripts/nemo\_ami\_benchmark\ directory. This tool runs NVIDIA's original NeMo Sortformer model on the AMI SDM dataset to establish a baseline for comparing against the Swift/CoreML implementation. The script supports streaming inference with configurable chunking parameters, handles RTTM ground truth files (including auto-download from pyannote), and outputs detailed diarization metrics such as DER, miss rate, and real-time factor.

_Scripts/nemo\_ami\benchmark · high confidence

New TTS documentation and benchmark data for Kokoro ANE v3, LuxTTS, and Chatterbox

The Documentation/TTS area now includes comprehensive reference material for the latest TTS backends. Benchmarks.md provides detailed latency and compute-routing measurements for Kokoro ANE v3 on Apple M5 Pro hardware, alongside methodology for historical corpus comparisons. KokoroAne.md documents the new v3 runtime, its hybrid compute policies, language variants (English, Japanese, Mandarin), and usage constraints. LuxTts.md covers the zero-shot voice-cloning backend, including platform-specific graph selection (GPU vs ANE) and G2P implementation. Chatterbox.md introduces the beta Multilingual and Nano backends, detailing their CoreML architectures, sampling defaults, and output capacity limits. Supporting measurement data (JSON) and a G2P comparison report are also added to substantiate these claims.

Documentation/TTS · high confidence

New benchmarking, profiling, and dataset materialization scripts

The Scripts directory now includes a suite of new tools to support evaluation and development workflows. A new Swift utility, ane\_profile.swift, reports the CoreML compute-unit plan (ANE, GPU, CPU) for .mlmodelc bundles. Several new benchmark scripts have been added: run\_benchmarks.py provides a unified Python entry point for ASR, VAD, and Diarization benchmarks with baseline comparisons; shell scripts (parakeet\_subset\_benchmark.sh, fleurs\_parakeet\_sub\_benchmark.sh, diarizer\_subset\_benchmark.sh) automate regression testing for Parakeet TDT, FLEURS multilingual, and diarization models respectively. Additionally, Python scripts (materialize\_earnings22\_longform.py, materialize\_alimeeting\_card\_audio.py, materialize\_notsofar\_card\_audio.py) and a dataset manifest (localvqe-dataset.json) help prepare specific evaluation datasets for long-form ASR and NVIDIA Nemotron 3 card-style benchmarks. Finally, verify\_localvqe\_benchmark.py and its test suite ensure the integrity of local VQE benchmark reports.

Scripts · high confidence

New macOS demo apps for Phonon and Pocket TTS small models

Two new SwiftUI demo applications are available for macOS 15+ to showcase Gradium's on-device speech synthesis capabilities. The PhononDemo app demonstrates the 31M Phonon model running on the Neural Engine, featuring a visual 'ring' that reacts to live audio spectrum and supports a variety of regular and character voices. The PocketTtsSmallDemo app provides an A/B comparison interface for the 18m, 7m, and 2m Pocket TTS models, allowing users to switch between CPU and Neural Engine execution modes and compare Core ML output against bundled PyTorch reference renders. Also in this area: New Bark Typewriter demo app for macOS; Fixes session race conditions and UI stability in the Pocket TTS demo; New macOS demo apps for LocalVQE acoustic enhancement and Phonon-2 speech recognition.

Examples · high confidence

Behavioural changes

Repository rebranded to FluidAudio with Swift 6.2+ opt-out for NeMo text processing

The project has been renamed from SeamlessAudioSwift to FluidAudio, including the Swift package name and all documentation. For Swift 6.2+ toolchains, the bundled NeMo text-normalization engine (\~8 MB) is now an opt-in trait, allowing ASR-only applications to exclude it by specifying \traits: \[\]\ in their package dependencies.

(repo-wide) · high confidence

Test coverage

2 commits adding/updating tests in Tests/FluidAudioTests/Speaker; 23 commits adding/updating tests in Tests/FluidAudioTests/ASR/Parakeet/SlidingWindow/CustomVocabulary; 3 commits adding/updating tests in Tests/FluidAudioTests/ASR/LFM3Audio; 3 commits adding/updating tests in Tests/FluidAudioTests/TTS/Paradee; Added CI test suite for FluidAudio diarizer and model components; Added LuxTTS test suite and fixtures; Added Phonon text normalizer golden fixture scripts; Added comprehensive test coverage for Diarizer components; Added comprehensive test coverage for KokoroAne TTS variants and internals; Added comprehensive test suite for Parakeet TDT sliding-window ASR; Added deterministic VAD test helper; Added regression and unit tests for ASR token deduplication, punctuation commit layer, and logits argmax; Added test coverage for CLI scoring determinism, AMI data handling, and LocalVQE enhancement; Added test coverage for FluidAudio core components; Added test coverage for VAD chunking, segmentation, and streaming logic; Added tests for CUA-S1-FORMS Core ML integration and output handling; Added tests for Chatterbox TTS alignment, assets, capacity, and tokenizers; Added tests for Cohere ASR configuration, compute units, long-form merging, and mask logic; Added tests for English TTS text normalization, initialism handling, and phoneme chunking; Added tests for G2P model directory resolution and registry behavior; Added tests for Inflect v2 CoreML backend components; Added tests for Nemotron 3 diarizer streaming, caching, and tensor layout; Added tests for Nemotron 3.5 Transcribe Lite core logic; Added tests for SlidingWindow ASR configuration, model loading, and streaming regression fixes; Added tests for diarizer extraction edge cases; Added unit tests for Canary ASR configuration, chunk merging, and segment packing; Added unit tests for Parakeet ASR configuration, constants, and decoding logic; Added unit tests for Parakeet streaming ASR components; Added unit tests for PocketTTS synthesis logic and infrastructure; Added unit tests for StyleTTS2 text processing and audio glue operations; Added unit tests for Supertonic-3 TTS internals; Added unit tests for TTS audio post-processing, text normalization, and language handling; Added unit tests for the offline diarizer module.

Dependencies

Initial release of FluidAudio Swift SDK and demo

This change introduces the FluidAudio Swift package (version 0.17.4) for macOS 14 and iOS 17, providing local audio AI capabilities including speaker diarization, voice-activity detection, and transcription via CoreML. The SDK includes a new \fluidaudiocli\ executable target and an optional \NemoTextProcessing\ binary dependency (v0.3.1) for byte-exact text normalization. A new \LocalVQEDemo\ example project is added to demonstrate usage, and the package is configured for Swift 6.0 with strict concurrency compliance.

(dependencies) · high confidence

Housekeeping

Added third-party license documentation for TTS components

New license files have been added to the ThirdPartyLicenses directory to document the provenance and usage terms for various text-to-speech resources. This includes attribution for the Japanese G2P frontend (covering unidic-lite, Misaki, and Cutlet), the Kokoro ANE-v3 toolkit (citing mattmireles/kokoro-coreml and hexgrad/Kokoro weights), the Spanish and French G2P frontends (referencing ipa-dict and CharsiuG2P), the LuxTTS English G2P resources (noting GPL-3.0 and CC-BY-SA-4.0 components from eSpeak NG and FrequencyWords), the NemoTextProcessing binary target (Apache-2.0), and the fastcluster and vbx libraries.

ThirdPartyLicenses · high confidence

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

Score

  • CAI 60 → 61 (+1.1)
  • Rubric changed (rubric-2026.10.5 → rubric-2026.11.4) — scores are not directly comparable.

Lenses

  • Code Health 59 → 67 (+7.7)
  • Architecture 90 → 98 (+8.0)
  • Maturity 68 → 70 (+2.5)
  • Readiness 55 → 54 (-0.7)
  • Security 61 → 61 (+0.0)
  • Domain Modelling 86 → 87 (+0.5)
  • Performance 71 → 69 (-2.8)

Resolved (125)

  • AudioConverter.convertBuffer (cognitive 16) (Sources/FluidAudio/Shared/AudioConverter.swift)
  • AudioSourceFactory.makeDiskBackedSource (cognitive 17) (Sources/FluidAudio/Shared/AudioSourceFactory.swift)
  • AudioSourceFactory.streamConvert (cognitive 20) (Sources/FluidAudio/Shared/AudioSourceFactory.swift)
  • Duplicated block (10 lines × 2) (Sources/FluidAudio/ASR/Parakeet/AsrTypes.swift)
  • Duplicated block (10 lines × 2) (Sources/FluidAudio/TTS/StyleTTS2/Pipeline/Synthesize/StyleTTS2Synthesizer.swift)
  • Duplicated block (10 lines × 2) (Sources/FluidAudio/TTS/StyleTTS2/StyleTTS2Error.swift)
  • Duplicated block (10 lines × 2) (Sources/FluidAudioCLI/Commands/ASR/Parakeet/SlidingWindow/CtcDecodeBenchmark.swift)
  • Duplicated block (10 lines × 3) (Sources/FluidAudio/TTS/Inflect/InflectError.swift)
  • Duplicated block (10 lines × 3) (Sources/FluidAudioCLI/Commands/ASR/Parakeet/SlidingWindow/AsrBenchmark.swift)
  • Duplicated block (10–11 lines × 3) (Sources/FluidAudioCLI/Commands/ASR/Cohere/CohereBenchmark.swift)
  • Duplicated block (10–13 lines × 2) (Sources/FluidAudio/ASR/Parakeet/Streaming/Nemotron/StreamingNemotronMultilingualAsrManager+Shared.swift)
  • Duplicated block (10–13 lines × 2) (Sources/FluidAudio/TTS/LuxTts/LuxTtsSynthesizer.swift)
  • Duplicated block (11 lines × 2) (Sources/FluidAudio/TTS/PocketTTS/Phonon/PhononCharTables.swift)
  • Duplicated block (11 lines × 2) (Sources/FluidAudioCLI/Commands/ASR/Cohere/CohereBenchmark.swift)
  • Duplicated block (12 lines × 2) (Sources/FluidAudio/ASR/Parakeet/SlidingWindow/CustomVocabulary/Rescorer/VocabularyRescorer.swift)
  • Duplicated block (13 lines × 2) (Sources/FluidAudioCLI/Commands/DiarizationBenchmark.swift)
  • Duplicated block (13 lines × 3) (Sources/FluidAudioCLI/Commands/ASR/Parakeet/Streaming/ParakeetEouCommand.swift)
  • Duplicated block (13–14 lines × 2) (Sources/FluidAudio/Diarizer/Sortformer/Offline/OfflineSortformerDiarizer.swift)
  • Duplicated block (13–14 lines × 3) (Sources/FluidAudio/TTS/StyleTTS2/Pipeline/Synthesize/StyleTTS2Synthesizer.swift)
  • Duplicated block (13–15 lines × 2) (Sources/FluidAudio/ASR/Parakeet/Streaming/EOU/StreamingEouAsrManager.swift)
  • …and 105 more

New (361)

  • AsrManager.unspacedKeptCopyCount (cognitive 24) (Sources/FluidAudio/ASR/Parakeet/SlidingWindow/TDT/AsrManager+Transcription.swift)
  • AsrManager.unspacedKeptCopyCount (cyclomatic 21) (Sources/FluidAudio/ASR/Parakeet/SlidingWindow/TDT/AsrManager+Transcription.swift)
  • AudioConverter.convertFile (cognitive 32) (Sources/FluidAudio/Shared/AudioConverter.swift)
  • AudioConverter.convertFile (cyclomatic 25) (Sources/FluidAudio/Shared/AudioConverter.swift)
  • AudioConverter.streamResample (cognitive 29) (Sources/FluidAudio/Shared/AudioConverter.swift)
  • AudioConverter.streamResample (cyclomatic 20) (Sources/FluidAudio/Shared/AudioConverter.swift)
  • Change coupling: ModelNames.swift ↔ KokoroAneSynthesizer+Types.swift (Sources/FluidAudio/ModelNames.swift)
  • Change coupling: SlidingWindowAsrManager.swift ↔ AsrManager+Transcription.swift (Sources/FluidAudio/ASR/Parakeet/SlidingWindow/SlidingWindowAsrManager.swift)
  • Change coupling: StreamingUnifiedAsrManager.swift ↔ UnifiedAsrManager.swift (Sources/FluidAudio/ASR/Parakeet/Unified/StreamingUnifiedAsrManager.swift)
  • ChunkProcessor.mergeChunks (cognitive 19) (Sources/FluidAudio/ASR/Parakeet/SlidingWindow/TDT/ChunkProcessor.swift)
  • ChunkProcessor.mergeChunks (cyclomatic 18) (Sources/FluidAudio/ASR/Parakeet/SlidingWindow/TDT/ChunkProcessor.swift)
  • CtcKeywordSpotter.makeLogProbs (cognitive 19) (Sources/FluidAudio/ASR/Parakeet/SlidingWindow/CustomVocabulary/WordSpotting/CtcKeywordSpotter+Inference.swift)
  • CtcKeywordSpotter.makeLogProbs (cyclomatic 18) (Sources/FluidAudio/ASR/Parakeet/SlidingWindow/CustomVocabulary/WordSpotting/CtcKeywordSpotter+Inference.swift)
  • Duplicated block (10 lines × 2) (Sources/FluidAudio/ASR/LFM3Audio/LFM3AudioEncoder.swift)
  • Duplicated block (10 lines × 2) (Sources/FluidAudio/ASR/Parakeet/AsrTypes.swift)
  • Duplicated block (10 lines × 2) (Sources/FluidAudio/ASR/Parakeet/Unified/StreamingUnifiedAsrManager.swift)
  • Duplicated block (10–11 lines × 3) (Sources/FluidAudioCLI/Commands/ASR/Cohere/CohereBenchmark.swift)
  • Duplicated block (11 lines × 3) (Sources/FluidAudioCLI/Commands/ASR/Parakeet/SlidingWindow/AsrBenchmark.swift)
  • Duplicated block (12 lines × 2) (Sources/FluidAudio/ASR/LFM3Audio/LFM3AudioTokenizer.swift)
  • Duplicated block (12 lines × 2) (Sources/FluidAudio/Diarizer/Sortformer/SortformerDiarizer.swift)
  • …and 341 more

Changes since last survey

  • 75 commits — 32 feature/other, 43 fixes

By area

  • Sources/FluidAudio — 50 commits
  • (root) — 8 commits
  • Documentation/ASR — 5 commits
  • .github/workflows — 3 commits
  • Documentation/Showcase.md — 3 commits
  • Tests/FluidAudioTests — 2 commits
  • Documentation/Benchmarks.md — 1 commit
  • Documentation/Diarization — 1 commit
  • Documentation/TTS — 1 commit
  • Sources/FluidAudioCLI — 1 commit

Notable commits

  • fix: fix(asr): Preserve names, possessives and surrounding words (#1041)
  • fix: fix(asr): flush pending tokens on an interior window that ends in a pause (#1111)
  • fix: fix(asr): recover dropped opening speech in long-form v3 decoding (#1056)
  • fix: fix(asr): seam-gap repair probes a short window around every inter-word gap (#1061)
  • fix: fix(asr): treat currency symbols as script-neutral in TokenLanguageFilter (#1120)
  • fix: fix(asr): wait for cancelled streaming recognition to exit (#1033)
  • fix: fix(asr/cohere): cut Japanese and Chinese long-form audio at pauses (#1146)
  • fix: fix(asr/nemotron-multilingual): encoder_v2 so M2/M3 Neural Engines compile the encoder (#1145)
  • fix: fix(asr/streaming): a re-spelled word at a window seam replaces the kept one (#1128)
  • fix: fix(asr/streaming): join a window that opens mid-word onto the previous text (#1097)
  • fix: fix(asr/tdtJa): drop a re-decode's copies of the kept tokens at a window seam (#1139) (#1143)
  • fix: fix(asr/tdtJa): long-form Japanese defaults to no mel context and probes only long seam gaps (#1098)
  • fix: fix(asr/tdtJa): streaming seam re-decodes a time tail, not the whole previous window (#1125)
  • fix: fix(asr/vocab): no context bias against a word TDT heard confidently (#1100)
  • fix: fix(asr/vocab): read CTC tokenizer next to the loaded CtcModels (#1114) (#1126)
  • fix: fix(asr/vocab): rescue similarity floors on by default, score the original as spoken (#1151) (#1152)
  • fix: fix(asr/vocab): score a lone stopword that is an exact alias (#1115) (#1127)
  • fix: fix(audio): give transcribe(url) and transcribe(samples) the same samples (#1134)
  • fix: fix(audio): resample long files without the 4 GiB buffer cap, converting files as they are read (#1021)
  • fix: fix(build): bump podspec to 0.18.0
  • …and 55 more

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

FluidInference/FluidAudio was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 14 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit 16780f04d2ece24f67d0fc00efccd0296491c1a8 — the exact code this score is about.
  • Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.65+preprod-3c33c1882dc1.