Transcriber list: revision 2

README.md

@@ -0,0 +1,16 @@
+---
+license: mit
+tags:
+- transcriber
+---
+
+# Transcriber: models and updates
+
+`transcriber.json` is the list the Transcriber app reads when it checks for
+updates: the speech, speaker, punctuation and clean-up models it offers (where
+each downloads from, and the checksums it must match), and the latest version
+of the app when its installer is published here (`installers/`).
+
+The app reads it only when its user asks (Check for updates) or has agreed to a
+daily check. Nothing about the user is sent beyond an ordinary download
+request.

transcriber.json

@@ -0,0 +1,223 @@
+{
+ "_about": "Read by the Transcriber app (Check for updates). models: as transcriber/catalog.json, used when revision is higher than the app's own. app: the latest version and its installer.",
+ "revision": 2,
+ "models": [
+  {
+   "id": "gpu-support",
+   "kind": "gpu",
+   "name": "GPU support (NVIDIA cuBLAS and cuDNN)",
+   "about": "Lets the speech models run on your NVIDIA graphics card, many times faster than on the processor. Shown only on computers with an NVIDIA card.",
+   "languages": "Any",
+   "license": "NVIDIA software licence",
+   "size_mb": 1373,
+   "download_mb": 1373,
+   "source": {
+    "pypi": true
+   }
+  },
+  {
+   "id": "large-v3",
+   "kind": "speech",
+   "name": "Best accuracy (Whisper large-v3)",
+   "about": "The most accurate general model. Best with an NVIDIA graphics card.",
+   "languages": "99 languages",
+   "license": "MIT",
+   "size_mb": 3091,
+   "cost": 1.0,
+   "vram_gb": 5,
+   "source": {
+    "repo": "Systran/faster-whisper-large-v3"
+   }
+  },
+  {
+   "id": "large-v3-turbo",
+   "kind": "speech",
+   "name": "Fast and accurate (Whisper large-v3 turbo)",
+   "about": "Nearly as accurate as large-v3 and several times faster. A good choice without a powerful graphics card.",
+   "languages": "99 languages",
+   "license": "MIT",
+   "size_mb": 1622,
+   "cost": 0.3,
+   "vram_gb": 3,
+   "source": {
+    "repo": "mobiuslabsgmbh/faster-whisper-large-v3-turbo"
+   }
+  },
+  {
+   "id": "distil-large-v3",
+   "kind": "speech",
+   "name": "Fast, English only (Distil-Whisper large-v3)",
+   "about": "Close to large-v3 accuracy on English, much faster. English recordings only.",
+   "languages": "English",
+   "license": "MIT",
+   "size_mb": 1516,
+   "cost": 0.3,
+   "vram_gb": 3,
+   "source": {
+    "repo": "Systran/faster-distil-whisper-large-v3"
+   }
+  },
+  {
+   "id": "medium",
+   "kind": "speech",
+   "name": "Medium (Whisper medium)",
+   "about": "Smaller and faster than large, a little less accurate.",
+   "languages": "99 languages",
+   "license": "MIT",
+   "size_mb": 1531,
+   "cost": 0.5,
+   "vram_gb": 2.5,
+   "source": {
+    "repo": "Systran/faster-whisper-medium"
+   }
+  },
+  {
+   "id": "small",
+   "kind": "speech",
+   "name": "Small (Whisper small)",
+   "about": "Quick on any computer; fine for clear recordings, weaker on accents and noise.",
+   "languages": "99 languages",
+   "license": "MIT",
+   "size_mb": 486,
+   "cost": 0.2,
+   "vram_gb": 1,
+   "source": {
+    "repo": "Systran/faster-whisper-small"
+   }
+  },
+  {
+   "id": "base",
+   "kind": "speech",
+   "name": "Base (Whisper base)",
+   "about": "Very fast, rough drafts only.",
+   "languages": "99 languages",
+   "license": "MIT",
+   "size_mb": 148,
+   "cost": 0.08,
+   "vram_gb": 0.5,
+   "source": {
+    "repo": "Systran/faster-whisper-base"
+   }
+  },
+  {
+   "id": "ivrit-large-v3-turbo",
+   "kind": "speech",
+   "name": "Hebrew, fast (ivrit.ai large-v3 turbo)",
+   "about": "Whisper trained further on Hebrew by ivrit.ai: much better Hebrew than the general models.",
+   "languages": "Hebrew (and English)",
+   "license": "Apache-2.0",
+   "size_mb": 1622,
+   "cost": 0.3,
+   "vram_gb": 3,
+   "source": {
+    "repo": "ivrit-ai/whisper-large-v3-turbo-ct2"
+   }
+  },
+  {
+   "id": "ivrit-large-v3",
+   "kind": "speech",
+   "name": "Hebrew, best (ivrit.ai large-v3)",
+   "about": "The most accurate Hebrew model from ivrit.ai. Best with an NVIDIA graphics card.",
+   "languages": "Hebrew (and English)",
+   "license": "Apache-2.0",
+   "size_mb": 3091,
+   "cost": 1.0,
+   "vram_gb": 5,
+   "source": {
+    "repo": "ivrit-ai/whisper-large-v3-ct2"
+   }
+  },
+  {
+   "id": "ecapa",
+   "kind": "voice",
+   "name": "Speaker recognition (SpeechBrain ECAPA, ONNX)",
+   "about": "Tells the voices apart. Needed for every transcription.",
+   "languages": "Any",
+   "license": "Apache-2.0",
+   "size_mb": 84,
+   "required": true,
+   "files": [
+    "ecapa-any-length.onnx"
+   ],
+   "source": {
+    "repo": "kenrunio/transcriber-speaker-onnx"
+   },
+   "sha256": {
+    "ecapa-any-length.onnx": "700b63064ee98981186f59f247d28e56f7f6ab265d8b50ebb450db5a8ce21dc7"
+   }
+  },
+  {
+   "id": "punctuation",
+   "kind": "punctuation",
+   "name": "Punctuation repair (fullstop multilang large, ONNX, 8-bit)",
+   "about": "Adds full stops and commas where Whisper leaves run-on text. Optional; English, German, French and Italian.",
+   "languages": "English, German, French, Italian",
+   "license": "MIT",
+   "size_mb": 581,
+   "files": [
+    "model_int8.onnx",
+    "tokenizer.json",
+    "labels.json"
+   ],
+   "older": [
+    "model.onnx",
+    "model.onnx.data",
+    "tokenizer.json",
+    "labels.json"
+   ],
+   "update_about": "A smaller version: 0.6 GB instead of 2.2 GB, a third faster, and the same punctuation on 99 words in 100. The old one is removed once it is in.",
+   "source": {
+    "repo": "kenrunio/transcriber-punctuation-onnx"
+   },
+   "sha256": {
+    "model_int8.onnx": "016c7f8ed917f564ca16792d838c9f7a1d91f4b567a29569108a862d2da65fb0",
+    "tokenizer.json": "8124d1990cd91a86da7d51b0bd10b7913377f48710901e17b7db7b6f8cd5aee6",
+    "labels.json": "a19c8dc4400604b04a4da3cc356d7c5709110a0ce6770bd0c067697db88fd5a7"
+   }
+  },
+  {
+   "id": "sounds",
+   "kind": "sounds",
+   "name": "Sound recognition (Audio Spectrogram Transformer, ONNX)",
+   "about": "For clean-up: tells coughs, throat clearing and sneezes from speech and other sounds. Optional.",
+   "languages": "Any",
+   "license": "BSD-3-Clause",
+   "size_mb": 91,
+   "files": [
+    "onnx/model_int8.onnx",
+    "config.json"
+   ],
+   "source": {
+    "repo": "onnx-community/ast-finetuned-audioset-10-10-0.4593-ONNX",
+    "revision": "2fe67046c486b8b8bc90a7b469239b2dec822f72"
+   },
+   "sha256": {
+    "onnx/model_int8.onnx": "6739e480a4faf4651f60d2238b0eb8cd58bead2a07ae1da5e2fbc72e3f7c98a6",
+    "config.json": "d4d347e95b00f4b1a0ea7eaf32bbbb2c246539ce1e64da89ffc0700c4a61ac8a"
+   }
+  },
+  {
+   "id": "align",
+   "kind": "align",
+   "name": "Word timing (wav2vec2 base, ONNX)",
+   "about": "For clean-up: finds exactly where each word is, so ums are cut without touching the words beside them. Optional.",
+   "languages": "English",
+   "license": "Apache-2.0",
+   "size_mb": 378,
+   "files": [
+    "onnx/model.onnx",
+    "vocab.json",
+    "config.json"
+   ],
+   "source": {
+    "repo": "onnx-community/wav2vec2-base-960h-ONNX",
+    "revision": "729c1a6730fb549c20a1c73a3d3f96f11020225e"
+   },
+   "sha256": {
+    "onnx/model.onnx": "00b7cc69516c1ab63c429e63a2b543e4d42bb77441ec5b98ee935de175b00de1",
+    "vocab.json": "4178db26b3c7570f6a47f14ac6a1c7b32950b8c2800fb097287e53776934f1c5",
+    "config.json": "15c7cf6378153bdcb33fce27780ab9aae37fd154fbb674cacc3347992055d323"
+   }
+  }
+ ]
+}