cstr/cohere-transcribe-03-2026-gguf overview

{
  "metadata": {},
  "card_data": {
    "license": "apache-2.0",
    "language": [
      "ar",
      "de",
      "el",
      "en",
      "es",
      "fr",
      "it",
      "ja",
      "ko",
      "nl",
      "pl",
      "pt",
      "vi",
      "zh"
    ],
    "pipeline_tag": "automatic-speech-recognition",
    "tags": [
      "audio",
      "speech-recognition",
      "transcription",
      "gguf",
      "conformer",
      "whisper.cpp"
    ],
    "base_model": "CohereLabs/cohere-transcribe-03-2026",
    "frontmatter": {
      "license": "apache-2.0",
      "language": [
        "ar",
        "de",
        "el",
        "en",
        "es",
        "fr",
        "it",
        "ja",
        "ko",
        "nl",
        "pl",
        "pt",
        "vi",
        "zh"
      ],
      "pipeline_tag": "automatic-speech-recognition",
      "tags": [
        "audio",
        "speech-recognition",
        "transcription",
        "gguf",
        "conformer",
        "whisper.cpp"
      ],
      "base_model": "CohereLabs/cohere-transcribe-03-2026"
    },
    "hero_image_url": "",
    "summary": "GGUF weights for **CohereLabs/cohere-transcribe-03-2026** — Cohere's open-source 2B-parameter ASR model, #1 on the Open ASR Leaderboard (avg WER 5.42, as of March 2026). This conversion enables high-performance CPU inference via **CrispASR** — a whisper.cpp-style C++ runtime for the Cohere Conformer-encoder / Transformer-decoder architecture. > **License**: Apache 2.0 (inherited from source model). See original model card for full terms. ---",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlicense: apache-2.0\nlanguage:\n- ar\n- de\n- el\n- en\n- es\n- fr\n- it\n- ja\n- ko\n- nl\n- pl\n- pt\n- vi\n- zh\npipeline_tag: automatic-speech-recognition\ntags:\n- audio\n- speech-recognition\n- transcription\n- gguf\n- conformer\n- whisper.cpp\nbase_model: CohereLabs/cohere-transcribe-03-2026\n---\n\n# cohere-transcribe-03-2026 — GGUF\n\nGGUF weights for **[CohereLabs/cohere-transcribe-03-2026](https://huggingface.co/CohereLabs/cohere-transcribe-03-2026)** — Cohere's open-source 2B-parameter ASR model, #1 on the [Open ASR Leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard) (avg WER 5.42, as of March 2026).\n\nThis conversion enables high-performance CPU inference via **[CrispASR](https://github.com/CrispStrobe/CrispASR)** — a whisper.cpp-style C++ runtime for the Cohere Conformer-encoder / Transformer-decoder architecture.\n\n> **License**: Apache 2.0 (inherited from source model). See [original model card](https://huggingface.co/CohereLabs/cohere-transcribe-03-2026) for full terms.\n\n---\n\n## Files\n\n| File | Size | Type | RTFx (8 threads) |\n|------|------|------|------------------|\n| `cohere-transcribe.gguf` | 3.85 GB | F16 | 0.80x |\n| `cohere-transcribe-q8_0.gguf` | 2.05 GB | Q8_0 | 1.03x |\n| `cohere-transcribe-q6_k.gguf` | 1.62 GB | Q6_K | 1.05x |\n| `cohere-transcribe-q5_1.gguf` | 1.45 GB | Q5_1 | 1.06x |\n| `cohere-transcribe-q5_0.gguf` | 1.38 GB | Q5_0 | 1.07x |\n| `cohere-transcribe-q4_k.gguf` | 1.21 GB | Q4_K | 1.08x |\n\n**RTFx** measured on `jfk.wav` (11s) using 8 CPU threads. Higher is faster. 1.0x means real-time.\n\n---\n\n## Quick Start\n\n### 1. Build CrispASR\n\n```bash\ngit clone -b ggml https://github.com/CrispStrobe/CrispASR\ncd CrispASR && mkdir build && cd build\ncmake .. -DCMAKE_BUILD_TYPE=Release\nmake -j$(nproc) cohere-main\n```\n\n### 2. Download a GGUF\n\n```bash\nhuggingface-cli download cstr/cohere-transcribe-03-2026-GGUF \\\n    cohere-transcribe-q4_k.gguf \\\n    --local-dir .\n```\n\n### 3. Transcribe\n\n```bash\n./bin/cohere-main \\\n    -m cohere-transcribe-q4_k.gguf \\\n    -f audio.wav \\\n    -l en \\\n    -t 8\n```\n\n---\n\n## Implementation Notes (Critical for Correctness)\n\n### Mel normalization\nPer-feature normalization uses **biased standard deviation** `std = sqrt(mean(diff²) + ε)`, matching the ONNX reference. Using the Bessel-corrected (unbiased) formula produces a `sqrt(T) ≈ 20×` larger denominator for T ≈ 417 frames and completely corrupts the encoder output.\n\n### Conformer Attention Scaling\nThe self-attention mechanism in the Conformer encoder **must** be scaled by `1/sqrt(head_dim)` before the softmax. Omitting this results in saturated attention scores and repetitive \"garbage\" output (e.g., \"what what what...\").\n\n### Encoder preprocessing\n1. **Pre-emphasis**: `y[n] = x[n] - 0.97·x[n-1]`\n2. **Center-pad**: `n_fft/2 = 256` samples on each side\n3. **STFT**: Hann window (length 400, zero-padded to 512), hop 160, rfft → power spectrum\n4. **Mel Filterbank**: 128 bins → log → per-feature norm (biased std)\n\n### Conv subsampling\n5 convolutions with 3 stride-2 steps reducing T_mel → T_enc ≈ T_mel/8:\n`conv0(ReLU) → conv2(DW) → conv3(PW,ReLU) → conv5(DW) → conv6(PW,ReLU) → linear(d=1280)`\n\n### Cross-Attention Pre-computation\nFor high performance, cross-attention Key and Value tensors are pre-computed once per utterance from the encoder output. In this implementation, these projections are performed as part of the encoder's GGML compute graph to leverage backend acceleration.\n\n### Decoder activation\nTransformer decoder FFN uses **ReLU** (not SiLU/Swish).\n\n---\n\n## Architecture\n\n| Component | Details |\n|-----------|---------|\n| **Encoder** | 48-layer Conformer, d=1280, heads=8, head_dim=160, ffn=5120, conv_kernel=9 |\n| **Decoder** | 8-layer causal Transformer, d=1024, heads=8, head_dim=128, ffn=4096, max_ctx=1024 |\n| **Vocab** | 16,384 SentencePiece tokens |\n| **Audio** | 16 kHz mono, 128 mel bins, n_fft=512, hop=160, win=400 |\n| **Parameters** | ~2B |\n\n---\n\n## Related\n\n- **Source model**: [CohereLabs/cohere-transcribe-03-2026](https://huggingface.co/CohereLabs/cohere-transcribe-03-2026)\n- **C++ runtime**: [CrispStrobe/CrispASR](https://github.com/CrispStrobe/CrispASR) — also hosts ports of `parakeet-tdt-0.6b-v3`, `canary-1b-v2`, and a universal multilingual forced aligner (`nfa-align`)\n- **Open ASR Leaderboard**: [hf-audio/open_asr_leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard)\n\n### Sister GGUF releases in the same family\n\n- [`cstr/cohere-transcribe-onnx-int4`](https://huggingface.co/cstr/cohere-transcribe-onnx-int4) — ONNX INT4 export of the same Cohere model\n- [`cstr/cohere-transcribe-onnx-int8`](https://huggingface.co/cstr/cohere-transcribe-onnx-int8) — ONNX INT8 export of the same Cohere model\n- [`cstr/parakeet-tdt-0.6b-v3-GGUF`](https://huggingface.co/cstr/parakeet-tdt-0.6b-v3-GGUF) — NVIDIA's 600M multilingual ASR with built-in word timestamps (faster, smaller, 25 EU languages)\n- [`cstr/parakeet_de_med-GGUF`](https://huggingface.co/cstr/parakeet_de_med-GGUF) — German medical PEFT fine-tune of the parakeet base\n- [`cstr/canary-1b-v2-GGUF`](https://huggingface.co/cstr/canary-1b-v2-GGUF) — NVIDIA's 978M multilingual ASR + speech translation with explicit `-sl/-tl` flags\n- [`cstr/canary-ctc-aligner-GGUF`](https://huggingface.co/cstr/canary-ctc-aligner-GGUF) — universal multilingual subword forced aligner (25 EU languages, ~78 ms MAE)\n\n### Use case → which runtime?\n\n| Need | Right tool |\n| --- | --- |\n| **Lowest English WER** (Open ASR Leaderboard #1) | **`cohere-main`** ← this repo |\n| Multilingual ASR + free word timestamps | `parakeet-main` ([cstr/parakeet-tdt-0.6b-v3-GGUF](https://huggingface.co/cstr/parakeet-tdt-0.6b-v3-GGUF)) |\n| Multilingual ASR + speech translation + explicit language control | `canary-main` ([cstr/canary-1b-v2-GGUF](https://huggingface.co/cstr/canary-1b-v2-GGUF)) |\n| Multilingual subword forced alignment of any transcript | `nfa-align` ([cstr/canary-ctc-aligner-GGUF](https://huggingface.co/cstr/canary-ctc-aligner-GGUF)) |\n| English-only character-level forced alignment (~30 ms MAE) | `cohere-align` (uses wav2vec2-large-xlsr-53-english) |\n",
    "related_quantizations": []
  },
  "tags": [
    "gguf",
    "audio",
    "speech-recognition",
    "transcription",
    "conformer",
    "whisper.cpp",
    "automatic-speech-recognition",
    "ar",
    "de",
    "el",
    "en",
    "es",
    "fr",
    "it",
    "ja",
    "ko",
    "nl",
    "pl",
    "pt",
    "vi",
    "zh",
    "base_model:CohereLabs/cohere-transcribe-03-2026",
    "base_model:quantized:CohereLabs/cohere-transcribe-03-2026",
    "license:apache-2.0",
    "region:us"
  ],
  "likes": 4,
  "downloads": 2958,
  "gated": false,
  "private": false,
  "last_modified": "2026-04-08T08:12:01.000Z",
  "created_at": "2026-03-31T08:22:53.000Z",
  "pipeline_tag": "automatic-speech-recognition",
  "library_name": ""
}

{
  "_id": "69cb845d9287084293a09c00",
  "id": "cstr/cohere-transcribe-03-2026-GGUF",
  "modelId": "cstr/cohere-transcribe-03-2026-GGUF",
  "sha": "2242638d5dfecc6f1dbe6c3a8713b97deb2e150f",
  "createdAt": "2026-03-31T08:22:53.000Z",
  "lastModified": "2026-04-08T08:12:01.000Z",
  "author": "cstr",
  "downloads": 2958,
  "likes": 4,
  "gated": false,
  "private": false,
  "pipeline_tag": "automatic-speech-recognition",
  "library_name": "",
  "siblings_count": 8
}