cstr/parakeet_de_med-gguf overview

{
  "metadata": {},
  "card_data": {
    "license": "cc-by-4.0",
    "language": [
      "de"
    ],
    "pipeline_tag": "automatic-speech-recognition",
    "tags": [
      "audio",
      "speech-recognition",
      "transcription",
      "ggml",
      "gguf",
      "parakeet",
      "tdt",
      "fastconformer",
      "german",
      "medical"
    ],
    "library_name": "ggml",
    "base_model": "johannhartmann/parakeet_de_med",
    "frontmatter": {
      "license": "cc-by-4.0",
      "language": [
        "de"
      ],
      "pipeline_tag": "automatic-speech-recognition",
      "tags": [
        "audio",
        "speech-recognition",
        "transcription",
        "ggml",
        "gguf",
        "parakeet",
        "tdt",
        "fastconformer",
        "german",
        "medical"
      ],
      "library_name": "ggml",
      "base_model": "johannhartmann/parakeet_de_med"
    },
    "hero_image_url": "",
    "summary": "GGUF / ggml conversions of johannhartmann/parakeet_de_med for use with the parakeet-main CLI from **CrispStrobe/CrispASR@parakeet**. parakeet_de_med is Johann Hartmann's PEFT decoder+joint fine-tune of nvidia/parakeet-tdt-0.6b-v3 specialised for **German medical documentation** (Arztbriefe). On the German medical test set it scores **3.28% WER** vs the base model's 11.73% — a 72% relative reduction. The fine-tune freezes the encoder and trains only the TDT decoder + joint head (18.1M out of 627M parameters, 2.89%). This means:",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlicense: cc-by-4.0\nlanguage:\n- de\npipeline_tag: automatic-speech-recognition\ntags:\n- audio\n- speech-recognition\n- transcription\n- ggml\n- gguf\n- parakeet\n- tdt\n- fastconformer\n- german\n- medical\nlibrary_name: ggml\nbase_model: johannhartmann/parakeet_de_med\n---\n\n# Parakeet-DE-Med — GGUF (ggml-quantised)\n\nGGUF / ggml conversions of [`johannhartmann/parakeet_de_med`](https://huggingface.co/johannhartmann/parakeet_de_med) for use with the `parakeet-main` CLI from **[CrispStrobe/CrispASR@parakeet](https://github.com/CrispStrobe/CrispASR/tree/parakeet)**.\n\n`parakeet_de_med` is Johann Hartmann's PEFT decoder+joint fine-tune of [`nvidia/parakeet-tdt-0.6b-v3`](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) specialised for **German medical documentation** (Arztbriefe). On the German medical test set it scores **3.28% WER** vs the base model's 11.73% — a 72% relative reduction.\n\nThe fine-tune freezes the encoder and trains only the TDT decoder + joint head (18.1M out of 627M parameters, 2.89%). This means:\n- The architecture is identical to `parakeet-tdt-0.6b-v3` (24-layer FastConformer encoder, 2-layer LSTM predictor, 8198-class TDT joint head)\n- The same GGUF converter, runtime, and CLI work as-is\n- The frozen encoder still uses the base model's auto-language detection — for clean German speech this works well, for accented or noisy audio you may want to fall back to a different runtime (see [comparison table](#which-runtime-should-i-use))\n\n## Files\n\n| File | Size | Notes |\n| --- | ---: | --- |\n| `parakeet_de_med.gguf`        | 1.26 GB | F16, full precision |\n| `parakeet_de_med-q8_0.gguf`   | 711 MB  | Q8_0, near-lossless |\n| `parakeet_de_med-q5_0.gguf`   | 516 MB  | Q5_0 |\n| `parakeet_de_med-q4_k.gguf`   | 467 MB  | **Q4_K — recommended default** |\n\nAll quantisations produce the same text on the German verification clip:\n> Leider zu spät. Leider zu spät.\n\n## Quick start\n\n```bash\n# 1. Build the runtime\ngit clone -b parakeet https://github.com/CrispStrobe/CrispASR\ncd CrispASR\ncmake -B build -DCMAKE_BUILD_TYPE=Release\ncmake --build build -j$(nproc) --target parakeet-main\n\n# 2. Download a quantisation\nhuggingface-cli download cstr/parakeet_de_med-GGUF \\\n    parakeet_de_med-q4_k.gguf --local-dir .\n\n# 3. Transcribe German audio\n./build/bin/parakeet-main \\\n    -m parakeet_de_med-q4_k.gguf \\\n    -f german_audio.wav -t 8\n```\n\nThe runtime is the same `parakeet-main` binary used for the base parakeet-tdt-0.6b-v3. All the usual flags work: `-vad-model` for Silero VAD slicing, `-ck N` for fixed chunking, `-ml N` for max chars per line, `-osrt`/`-ovtt`/`-ot` for subtitle output, `-v` for per-token timestamps via the TDT duration head.\n\n## Word-level timestamps\n\nLike the base parakeet model, this fine-tune emits TDT durations as part of decoding, so word-level timestamps come for free at one encoder frame = **80 ms** granularity. No separate forced alignment model needed:\n\n```\n$ ./build/bin/parakeet-main -m parakeet_de_med-q4_k.gguf -f german.wav -t 8 -v\n[ 0.32s →  0.64s]  Der\n[ 0.64s →  1.04s]  Patient\n[ 1.04s →  1.32s]  klagt\n[ 1.32s →  1.92s]  über\n...\n```\n\n## Which runtime should I use?\n\nFor German speech specifically:\n\n| Use case | Right tool |\n| --- | --- |\n| **German medical documentation** | **`parakeet_de_med-q4_k.gguf`** ← this repo |\n| General German ASR with explicit language control | `canary-1b-v2-q4_k.gguf` (`-sl de -tl de`) |\n| German → English translation | `canary-1b-v2-q4_k.gguf` (`-sl de -tl en`) |\n| General multilingual ASR (auto-detect) | `parakeet-tdt-0.6b-v3-q4_k.gguf` |\n| Lowest English WER | `cohere-transcribe-q4_k.gguf` |\n\n## Architecture (inherited from base)\n\n| Component | Details |\n| --- | --- |\n| Encoder       | 24-layer FastConformer (frozen), d=1024, 8 heads, head_dim=128, FFN=4096, conv kernel=9 |\n| Subsampling   | Conv2d dw_striding stack, 8× temporal (100 → 12.5 fps) |\n| Predictor     | 2-layer LSTM, hidden 640, embed 8193 × 640 (**fine-tuned**) |\n| Joint head    | enc(1024 → 640) + pred(640 → 640) → ReLU → linear(640 → 8198) (**fine-tuned**) |\n| Vocab         | 8192 SentencePiece tokens (multilingual, but generation biased toward German medical) |\n| Audio         | 16 kHz mono, 128 mel bins, n_fft=512, hop=160, win=400 |\n| Parameters    | 627M total, 18.1M trained (2.89%) |\n\n## Attribution\n\n- **Base model:** [`nvidia/parakeet-tdt-0.6b-v3`](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) (CC-BY-4.0). NVIDIA NeMo team.\n- **Fine-tune:** [`johannhartmann/parakeet_de_med`](https://huggingface.co/johannhartmann/parakeet_de_med) (CC-BY-4.0). Johann Hartmann. Trained on 976 German medical documentation samples for 5 epochs with PEFT decoder+joint strategy.\n- **GGUF conversion + ggml runtime:** [`CrispStrobe/CrispASR@parakeet`](https://github.com/CrispStrobe/CrispASR/tree/parakeet).\n\n## Related\n\n- C++ runtime: **[CrispStrobe/CrispASR@parakeet](https://github.com/CrispStrobe/CrispASR/tree/parakeet)**\n- Base multilingual model (auto-detect): [`cstr/parakeet-tdt-0.6b-v3-GGUF`](https://huggingface.co/cstr/parakeet-tdt-0.6b-v3-GGUF)\n- Encoder–decoder companion (canary, with explicit language control + speech translation): [`cstr/canary-1b-v2-GGUF`](https://huggingface.co/cstr/canary-1b-v2-GGUF)\n- Cohere Transcribe (lowest English WER): [`cstr/cohere-transcribe-03-2026-GGUF`](https://huggingface.co/cstr/cohere-transcribe-03-2026-GGUF)\n\n## License\n\nCC-BY-4.0, inherited from both the base model and the fine-tune. Use of these GGUF files must comply with the CC-BY-4.0 license including attribution to NVIDIA NeMo team and Johann Hartmann.\n",
    "related_quantizations": []
  },
  "tags": [
    "ggml",
    "gguf",
    "audio",
    "speech-recognition",
    "transcription",
    "parakeet",
    "tdt",
    "fastconformer",
    "german",
    "medical",
    "automatic-speech-recognition",
    "de",
    "base_model:johannhartmann/parakeet_de_med",
    "base_model:quantized:johannhartmann/parakeet_de_med",
    "license:cc-by-4.0",
    "region:us"
  ],
  "likes": 0,
  "downloads": 197,
  "gated": false,
  "private": false,
  "last_modified": "2026-04-08T06:01:38.000Z",
  "created_at": "2026-04-08T05:05:51.000Z",
  "pipeline_tag": "automatic-speech-recognition",
  "library_name": "ggml"
}

{
  "_id": "69d5e22f8b033a517a753033",
  "id": "cstr/parakeet_de_med-GGUF",
  "modelId": "cstr/parakeet_de_med-GGUF",
  "sha": "8a785ea4c0bb284ce2a6c039129648cf7432c1e7",
  "createdAt": "2026-04-08T05:05:51.000Z",
  "lastModified": "2026-04-08T06:01:38.000Z",
  "author": "cstr",
  "downloads": 197,
  "likes": 0,
  "gated": false,
  "private": false,
  "pipeline_tag": "automatic-speech-recognition",
  "library_name": "ggml",
  "siblings_count": 6
}