cohere-transcribe-03-2026 — GGUF
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 crispasr-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.
Files
| File | Size | Type | RTFx (8 threads) |
|---|
cohere-transcribe.gguf | 3.85 GB | F16 | 0.80x |
cohere-transcribe-q8_0.gguf | 2.05 GB | Q8_0 | 1.03x |
cohere-transcribe-q6_k.gguf | 1.62 GB | Q6_K | 1.05x |
cohere-transcribe-q5_1.gguf | 1.45 GB | Q5_1 | 1.06x |
cohere-transcribe-q5_0.gguf | 1.38 GB | Q5_0 | 1.07x |
cohere-transcribe-q4_k.gguf | 1.21 GB | Q4_K | 1.08x |
RTFx measured on jfk.wav (11s) using 8 CPU threads. Higher is faster. 1.0x means real-time.
Quick Start
1. Build CrispASR
1git clone --recursive https://github.com/CrispStrobe/CrispASR
2cd CrispASR && mkdir build && cd build
3cmake .. -DCMAKE_BUILD_TYPE=Release
4cmake --build . -j $(nproc) --target crispasr-cli
2. Download a GGUF
1huggingface-cli download cstr/cohere-transcribe-03-2026-GGUF \
2 cohere-transcribe-q4_k.gguf \
3 --local-dir .
3. Transcribe
1./bin/crispasr --backend cohere \
2 -m cohere-transcribe-q4_k.gguf \
3 -f audio.wav \
4 -l en \
5 -t 8
Implementation Notes (Critical for Correctness)
The language whitelist is metadata, not vocab
config.json lists 14 supported languages (en fr de es it pt nl pl el ar ja zh vi ko), but the tokenizer carries all 183 ISO-639-1 <|xx|> tokens. So
<|ru|> decodes without error, and the model answers a wrong language
fluently rather than failing — there is no runtime signal at all. The
whitelist therefore has to travel with the weights: these GGUFs carry it as
cohere_transcribe.supported_languages, and CrispASR substitutes loudly when
-l names something outside it. For a GGUF converted before that key existed,
declare it with CRISPASR_COHERE_LANGS=en,fr,de,….
This matters most because -l auto runs an external detector that knows 99
languages against a model that accepts 14.
Mel normalization
Per-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.
Conformer Attention Scaling
The 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...").
Encoder preprocessing
- Pre-emphasis:
y[n] = x[n] - 0.97·x[n-1]
- Center-pad:
n_fft/2 = 256 samples on each side
- STFT: Hann window (length 400, zero-padded to 512), hop 160, rfft → power spectrum
- Mel Filterbank: 128 bins → log → per-feature norm (biased std)
Conv subsampling
5 convolutions with 3 stride-2 steps reducing T_mel → T_enc ≈ T_mel/8:
conv0(ReLU) → conv2(DW) → conv3(PW,ReLU) → conv5(DW) → conv6(PW,ReLU) → linear(d=1280)
Cross-Attention Pre-computation
For 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.
Decoder activation
Transformer decoder FFN uses ReLU (not SiLU/Swish).
Architecture
| Component | Details |
|---|
| Encoder | 48-layer Conformer, d=1280, heads=8, head_dim=160, ffn=5120, conv_kernel=9 |
| Decoder | 8-layer causal Transformer, d=1024, heads=8, head_dim=128, ffn=4096, max_ctx=1024 |
| Vocab | 16,384 SentencePiece tokens |
| Audio | 16 kHz mono, 128 mel bins, n_fft=512, hop=160, win=400 |
| Parameters | ~2B |
Related
- Source model: CohereLabs/cohere-transcribe-03-2026
- C++ runtime: CrispStrobe/CrispASR — also hosts ports of
parakeet-tdt-0.6b-v3, canary-1b-v2, and a universal multilingual forced aligner (nfa-align)
- Open ASR Leaderboard: hf-audio/open_asr_leaderboard
Sister GGUF releases in the same family
cstr/cohere-transcribe-onnx-int4 — ONNX INT4 export of the same Cohere modelcstr/cohere-transcribe-onnx-int8 — ONNX INT8 export of the same Cohere modelcstr/parakeet-tdt-0.6b-v3-GGUF — NVIDIA's 600M multilingual ASR with built-in word timestamps (faster, smaller, 25 EU languages)cstr/parakeet_de_med-GGUF — German medical PEFT fine-tune of the parakeet basecstr/canary-1b-v2-GGUF — NVIDIA's 978M multilingual ASR + speech translation with explicit -sl/-tl flagscstr/canary-ctc-aligner-GGUF — universal multilingual subword forced aligner (25 EU languages, ~78 ms MAE)
Use case → which runtime?
| Need | Right tool |
|---|
| Lowest English WER (Open ASR Leaderboard #1) | --backend cohere ← this repo |
| Multilingual ASR + free word timestamps | parakeet-main (cstr/parakeet-tdt-0.6b-v3-GGUF) |
| Multilingual ASR + speech translation + explicit language control | canary-main (cstr/canary-1b-v2-GGUF) |
| Multilingual subword forced alignment of any transcript | nfa-align (cstr/canary-ctc-aligner-GGUF) |
| English-only character-level forced alignment (~30 ms MAE) | cohere-align (uses wav2vec2-large-xlsr-53-english) |
Provenance and EU AI Act Art. 53 note
- Upstream model: CohereLabs/cohere-transcribe-03-2026 — published by
CohereLabs.
- Upstream licence:
apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
- What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
- Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
- Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.
