Predicts **word-level timestamps** for audio+text pairs in a single non-autoregressive forward pass. Unlike ASR (autoregressive, token-by-token), the forced aligner runs the **entire sequence in one forward pass** through the decoder. The classify head predicts a timestamp class (0–4999) at each ` ` token position, which maps to time via `class_index × 80ms`. Text decoder (attention projections, MLP, embeddings) quantized to 4-bit using group quantization (group_size=64). Audio encoder and classify head kept as float16 for accuracy.
MLX-compatible weights for Silero VAD v5, converted from the official JIT model. Silero VAD v5 is a lightweight (~309K params) voice activity detection model that processes 512-sample chunks (32ms @ 16kHz) with sub-millisecond latency. It outputs a speech probability between 0 and 1 for each chunk, with LSTM state carried across chunks for streaming operation. **Architecture:** STFT → 4×Conv1d+ReLU encoder → LSTM(128) → Conv1d decoder → sigmoid
Measured with `speech-swift` release tests on Apple Silicon using a 20 s 16 kHz speech fixture, 625 streaming chunks. The exported safetensors were also checked tensor-by-tensor against the upstream v6.2.1 JIT state dict after conversion; the maximum absolute tensor difference was 0.
**First-party CoreML export** of Supertonic-3's four non-autoregressive flow-matching graphs, for on-device iOS / Apple Neural Engine. Built by our own pipeline (`speech-models/stmodels`): weights lifted from the `Supertone/supertonic-3` ONNX initializers → PyTorch `nn.Module` → `coremltools` (`mlprogram`, FP32, iOS18+). Text/duration use **fixed T=128** (relpos attention has T-dependent pad widths — pad/segment text to 128); vocoder + vector-estimator use a dynamic latent-length `RangeDim`. The host runs the flow-matching ODE loop (`vector_estimator` ×`total_steps`) — the graphs contain no control flow. Assets to drive them: `tts.json`, `unicode_indexer.json` (G2P-free tokenizer table), `voice_styles/*.json`. > FP32 = parity reference. For **ANE residency**, use the mixed-precision > `Supertonic-3-CoreML-FP16` — vocoder + duration FP16, > text-encoder + vector-estimator FP32; measured transparent at **47–51 dB** mag-STFT SNR.
MLX-compatible weights for WeSpeaker ResNet34-LM, converted from the pyannote speaker embedding model with BatchNorm fused into Conv2d. WeSpeaker ResNet34-LM is a speaker embedding model (~6.6M params) that produces 256-dimensional L2-normalized speaker embeddings from audio. Trained on VoxCeleb for speaker verification and diarization. Converts the original pyannote/wespeaker-voxceleb-resnet34-LM checkpoint using a custom unpickler (no pyannote.audio dependency required). Key transformations:
We applied a version of our fine-tune data set onto Japanese Stable LM Base Gamma 7B and it performed pretty well, just sharing since it might be of interest.
Conversão sem fusões do modelo `dominguesm/legal-bert-ner-base-cased-ptbr` para uso com Transformers.js em WebGPU e WASM. O tokenizer diferencia maiúsculas e minúsculas e tem `model_max_length=512`. Além de PESSOA, ORGANIZACAO e LOCAL, o modelo produz TEMPO, LEGISLACAO e JURISPRUDENCIA. O benchmark da Epryx compara somente as três primeiras classes. Em oito frases jurídicas (129 tokens válidos), o FP16 preservou 129/129 predições `argmax` do ONNX FP32. O erro absoluto médio nos logits foi 0,001120 e o máximo foi 0,007393.
