Whisper large-v3 — Hush Core ML conversion
Apache-2.0Mac Speech-to-text (the Mac app's default transcription model)
Whisper large-v3 by OpenAI (Radford et al., "Robust Speech Recognition via Large-Scale Weak Supervision", 2022).
- Changes
- Converted by Hush from OpenAI's original weights to Core ML (WhisperKit format) with Argmax's MIT-licensed whisperkittools. No fine-tuning, quantization or other weight changes. The tokenizer files are redistributed unchanged.
- Source
- huggingface.co/openai/whisper-large-v3 @ 06f233fe06e7
- Delivery
- Hosted by Hush at downloads.hushmeetings.com.
Whisper large-v3-turbo — Hush Core ML conversion
MITMac Speech-to-text (the "Faster" choice in Settings on Mac)
Whisper large-v3-turbo by OpenAI. Copyright (c) 2022 OpenAI.
- Changes
- Converted by Hush from OpenAI's original weights to Core ML (WhisperKit format) with whisperkittools. No weight changes.
- Source
- huggingface.co/openai/whisper-large-v3-turbo @ 41f01f3fe87f
- Delivery
- Hosted by Hush at downloads.hushmeetings.com; downloaded only when you choose Faster.
Whisper large-v3-turbo — Hush CTranslate2 conversion
MITWindows Speech-to-text (the Windows app's default, "Most accurate")
Whisper large-v3-turbo by OpenAI. Copyright (c) 2022 OpenAI.
- Changes
- Converted by Hush from OpenAI's original weights to CTranslate2 format with 8-bit integer (int8) quantization. tokenizer.json and preprocessor_config.json are copied unchanged.
- Source
- huggingface.co/openai/whisper-large-v3-turbo @ 41f01f3fe87f
- Delivery
- Included in the Windows installer; also hosted by Hush at downloads.hushmeetings.com.
Whisper small.en — Hush CTranslate2 conversion
Apache-2.0Windows Speech-to-text (Windows "Faster, less accurate" option)
Whisper small.en by OpenAI (Radford et al., 2022).
- Changes
- Converted by Hush from OpenAI's original weights to CTranslate2 format at 16-bit floating point (float16). tokenizer.json and preprocessor_config.json are copied unchanged.
- Source
- huggingface.co/openai/whisper-small.en @ e8727524f962
- Delivery
- Included in the Windows installer; also hosted by Hush at downloads.hushmeetings.com.
Mac Live transcript while recording
Whisper large-v3-turbo by OpenAI (weights licensed MIT, Copyright (c) 2022 OpenAI); MLX conversion published by the mlx-community organization on Hugging Face, which states no license for its conversion.
- Changes
- Not modified by Hush.
- Source
- huggingface.co/mlx-community/whisper-large-v3-turbo @ a4aaeec0636e
- Delivery
- Downloaded from Hugging Face during setup.
Mac Backup transcription engine (used only when the faster engine is unavailable)
Whisper large-v3 by OpenAI, whose weights are licensed Apache-2.0; MLX conversion published by the mlx-community organization on Hugging Face, whose model card states MIT for the conversion. Apache-2.0 continues to govern OpenAI's weights.
- Changes
- Not modified by Hush.
- Source
- huggingface.co/mlx-community/whisper-large-v3-mlx @ 49e6aa286ad6
- Delivery
- Downloaded from Hugging Face only if needed.
Whisper large-v3 — Argmax WhisperKit conversion (previous default)
NOASSERTIONMac Speech-to-text (only on installs that still have it, or when selected manually)
Whisper large-v3 by OpenAI (weights licensed Apache-2.0); Core ML conversion by Argmax, Inc. (argmaxinc/whisperkit-coreml, variant openai_whisper-large-v3_turbo). Argmax states no license for its conversion in that repository.
- Changes
- Not modified by Hush.
- Source
- huggingface.co/argmaxinc/whisperkit-coreml @ 97a5bf9bbc74
- Delivery
- No longer downloaded by default. Earlier installs keep their copy until Hush's own conversion is in place.
SpeakerKit speaker models — Argmax Core ML
CC-BY-4.0Mac Telling speakers apart (segmentation, voice embeddings, clustering)
"SpeakerKit Core ML" models by Argmax, Inc., licensed under CC BY 4.0.
- Changes
- Not modified. Hush redistributes the 20 files it uses byte-for-byte and checks every file's SHA-256 before use.
- Source
- huggingface.co/argmaxinc/speakerkit-coreml @ 556fc52a1332
- Delivery
- Hosted by Hush at downloads.hushmeetings.com.
Built on
- pyannote segmentation-3.0 MITCopyright (c) 2023 CNRS Converted to Core ML by Argmax. huggingface.co/pyannote/segmentation-3.0
- WeSpeaker VoxCeleb ResNet34-LM CC-BY-4.0WeSpeaker project contributors; Wang et al., "WeSpeaker: A Research and Production Oriented Speaker Embedding Learning Toolkit"; Zeinali et al., "BUT System Description to VoxCeleb Speaker Recognition Challenge 2019". Trained on VoxCeleb2. Packaged by pyannote; converted to Core ML and quantized by Argmax. huggingface.co/Wespeaker/wespeaker-voxceleb-resnet34-LM
- VBx PLDA speaker-clustering model Apache-2.0BUT Speech@FIT, Brno University of Technology. Packaged as SpeakerKit's PLDA projector by Argmax. github.com/BUTSpeechFIT/VBx
Argmax does not endorse Hush. No warranties are given. Neither WeSpeaker, pyannote, the model authors nor the VoxCeleb creators endorse Hush.
pyannote segmentation-3.0 — ONNX conversion
MITWindows Telling speakers apart (finding speaker turns)
pyannote segmentation-3.0 by CNRS / pyannote; ONNX conversion by the onnx-community organization on Hugging Face (MIT). Copyright (c) 2023 CNRS.
- Changes
- Not modified by Hush.
- Source
- huggingface.co/onnx-community/pyannote-segmentation-3.0 @ 733a93b6473d
- Delivery
- Included in the Windows installer.
SpeechBrain ECAPA-TDNN speaker embedder — Hush ONNX export
Apache-2.0Windows Recognising voices (speaker embeddings)
spkrec-ecapa-voxceleb by the SpeechBrain project (Ravanelli et al., "SpeechBrain: A General-Purpose Speech Toolkit", 2021); ECAPA-TDNN by Desplanques, Thienpondt and Demuynck (Interspeech 2020). Trained on VoxCeleb 1 and 2.
- Changes
- Exported by Hush from the original weights to ONNX (opset 17). Hush reimplements SpeechBrain's audio features in its own code.
- Source
- huggingface.co/speechbrain/spkrec-ecapa-voxceleb @ 0f99f2d0ebe8
- Delivery
- Included in the Windows installer.
Gemma 4 (E4B, 12B, 26B-A4B, 31B) — GGUF builds
Apache-2.0Mac On-device summaries, action items and answers
Gemma 4 by Google DeepMind; 4-bit (Q4_K_M) GGUF quantizations by bartowski. The tier chosen in Settings decides which one is downloaded.
- Changes
- Not modified by Hush.
- Source
- ai.google.dev/gemma
- Delivery
- Downloaded from Hugging Face during setup.
Built on
Google does not endorse Hush. Gemma is a trademark of Google LLC.
Windows On-device summaries, action items and answers
Gemma 4 by Google DeepMind, as packaged by Ollama (model gemma4:e4b).
- Changes
- Not modified by Hush.
- Source
- ollama.com/library/gemma4
- Delivery
- Downloaded by Ollama, which you install separately, during setup.
Google does not endorse Hush. Gemma is a trademark of Google LLC.
Mac Search across meetings
nomic-embed-text-v1.5 by Nomic AI.
- Changes
- Not modified by Hush.
- Source
- huggingface.co/nomic-ai/nomic-embed-text-v1.5-GGUF @ 0188c9bf4097
- Delivery
- Downloaded from Hugging Face during setup.