tag:github.com,2008:https://github.com/Otosaku/NeMoFeatureExtractor-iOS/releases 2026-02-06T09:16:47Z tag:github.com,2008:Repository/1151224479/1.0.5 2026-02-06T09:32:04Z <p>v1.0.5: cleanup debug logging, fix resource loading</p> <p>- Simplify mel_filterbank.bin loading (supports both .copy and .process) <br />- Remove debug print statements <br />- Update README with correct org URL and version</p> <p>🤖 Generated with [Claude Code](<a href="https://claude.com/claude-code">https://claude.com/claude-code</a>)</p> <p>Co-Authored-By: Claude &lt;noreply@anthropic.com&gt;</p> otosaku-ai tag:github.com,2008:Repository/1151224479/1.0.4 2026-02-06T09:01:26Z <p>debug: add logging for resource loading</p> <p>🤖 Generated with [Claude Code](<a href="https://claude.com/claude-code">https://claude.com/claude-code</a>)</p> <p>Co-Authored-By: Claude &lt;noreply@anthropic.com&gt;</p> otosaku-ai tag:github.com,2008:Repository/1151224479/1.0.3 2026-02-06T08:49:19Z <p>fix: use .process instead of .copy for resources</p> <p>Fixes code signing issues on iOS simulator where .copy creates <br />an unrecognized bundle format.</p> <p>🤖 Generated with [Claude Code](<a href="https://claude.com/claude-code">https://claude.com/claude-code</a>)</p> <p>Co-Authored-By: Claude &lt;noreply@anthropic.com&gt;</p> otosaku-ai tag:github.com,2008:Repository/1151224479/v1.0.2 2026-02-06T08:27:29Z <p>Fix frame count formula based on NeMo source code</p> <p>Analyzed NeMo's FilterbankFeatures.get_seq_len() and forward() methods.</p> <p>Correct formula: <br />1. STFT frames = 1 + audio_length // hop_length (torch.stft with center=True) <br />2. Output frames = round_up(stft_frames, pad_to) if pad_to &gt; 0</p> <p>Previous formula (n + windowSize) / hopLength was incorrect.</p> <p>Now all model configs match NeMo exactly: <br />- VAD: [80, 52] == [80, 52] (pad_to=2) <br />- Speaker: [80, 64] == [80, 64] (pad_to=16) <br />- ASR: [80, 51] == [80, 51] (pad_to=0)</p> <p>Also improved tests: <br />- Added strict frame count equality checks <br />- Updated tolerances to 1e-4 max diff, 1e-5 avg diff</p> <p>🤖 Generated with [Claude Code](<a href="https://claude.com/claude-code">https://claude.com/claude-code</a>)</p> <p>Co-Authored-By: Claude &lt;noreply@anthropic.com&gt;</p> otosaku-ai tag:github.com,2008:Repository/1151224479/v1.0.1 2026-02-06T08:18:00Z <p>Fix STFT frame count formula to match NeMo exactly</p> <p>Changed frame count calculation for center padding from: <br /> sampleCount / hopLength + 1 (gave 51 frames) <br />to: <br /> (sampleCount + windowSize) / hopLength (gives 52 frames)</p> <p>This matches NeMo's AudioToMelSpectrogramPreprocessor output exactly. <br />Also updated output frame calculation to use nFrames instead of validFrames.</p> <p>Test results now show exact shape match: <br />- Swift mel shape: [80, 52] <br />- NeMo mel shape: [80, 52] <br />- Max diff: 5.4e-05</p> <p>🤖 Generated with [Claude Code](<a href="https://claude.com/claude-code">https://claude.com/claude-code</a>)</p> <p>Co-Authored-By: Claude &lt;noreply@anthropic.com&gt;</p> otosaku-ai tag:github.com,2008:Repository/1151224479/v1.0.0 2026-02-06T07:47:49Z <p>Initial release v1.0.0 - NeMoFeatureExtractor for iOS</p> <p>Swift library for extracting mel-spectrogram features compatible with <br />NVIDIA NeMo speech models. Features:</p> <p>- Exact compatibility with NeMo's feature extraction pipeline <br />- Supports VAD (MarbleNet), Speaker (TitaNet), and ASR models <br />- High performance using Apple's Accelerate framework (vDSP) <br />- Pre-computed mel filterbank from NeMo for maximum accuracy <br />- Output as [[Float]] or MLMultiArray for CoreML inference <br />- Tested against NeMo Python reference (max diff &lt; 6e-05)</p> <p>🤖 Generated with [Claude Code](<a href="https://claude.com/claude-code">https://claude.com/claude-code</a>)</p> <p>Co-Authored-By: Claude &lt;noreply@anthropic.com&gt;</p> otosaku-ai