cstr / canary-1b-v2-GGUF

Canary 1B v2 — GGUF (ggml-quantised)

GGUF / ggml conversions of nvidia/canary-1b-v2 for use with the canary-main CLI from CrispStrobe/cohere-whisper.cpp@parakeet .

Canary 1B v2 is NVIDIA's 978 M-parameter multilingual ASR + speech translation model:

• 25 European languages with explicit source_lang / target_lang task tokens (no auto-detect ambiguity)
• Speech translation in both directions: X→English (24 languages) and English→X (24 languages)
• 7.15% avg WER on the HuggingFace Open ASR Leaderboard (English) — competitive with Whisper-large-v3 at 1/1.6× the size
• CC-BY-4.0 licence

This is the encoder–decoder companion to cstr/parakeet-tdt-0.6b-v3-GGUF , which is the same FastConformer encoder family but with a TDT decoder for ASR-only. Both share the runtime and were ported in the same fork.

Files

Quick Start

# 1. Build the runtime
git clone -b parakeet https://github.com/CrispStrobe/cohere-whisper.cpp
cd cohere-whisper.cpp
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc) --target canary-main

# 2. Download a quantisation
huggingface-cli download cstr/canary-1b-v2-GGUF \
    canary-1b-v2-q4_k.gguf --local-dir .

# 3. ASR (English → English)
./build/bin/canary-main \
    -m canary-1b-v2-q4_k.gguf \
    -f your-audio.wav \
    -sl en -tl en -t 8

# 4. ASR (German → German)
./build/bin/canary-main -m canary-1b-v2-q4_k.gguf \
    -f german_audio.wav -sl de -tl de

# 5. Speech translation (German → English)
./build/bin/canary-main -m canary-1b-v2-q4_k.gguf \
    -f german_audio.wav -sl de -tl en

Verified end-to-end output

English ASR ( samples/jfk.wav , 11 s):

And so, my fellow Americans, ask not what your country can do for you, ask what you can do for your country.

German ASR (Wikimedia Commons Amardeo_Sarma_voice_-_de.ogg , 91 s):

Ich heiße Amadeus Scharma. Ich bin 1955 in Kassel in Deutschland geboren, weitgehend in Indien aufgewachsen. Ich hatte meine Ausbildung in Ingenieurwissenschaften in Neu-Delhi und dann später auch in Darmstadt, an der Technischen Universität von Darmstadt. ...

Speech translation DE → EN (same clip):

My name is Amadeo Sharma. I was born in Kassel in Germany in 1955, and I grew up largely in India. I had my education in engineering in New Delhi and then later also in Darmstadt, at the Technical University of Darmstadt. ...

Why explicit language tokens?

Auto-detect language ID can misfire on accented or noisy speech. We tested parakeet (which has no -l flag and relies on auto-detect) on the same German clips and it picked Russian for Angela Merkel and code-switched into English on Sarma's recording. Canary's -sl LANG removes that whole class of failures by telling the decoder explicitly what language to expect — see test_german.md in the runtime repo.

Supported languages

bg cs da de el en es et fi fr hr hu it lt lv mt nl pl pt ro ru sk sl sv uk (25 European languages).

For each pair ( source_lang , target_lang ):

• sl == tl → ASR
• sl != tl → speech translation

Translation supports any pair from 24 non-English languages → English, and English → any of 24 non-English languages.

Architecture

The mel filterbank and Hann window are baked into the GGUF ( preprocessor.fb and preprocessor.window ), so no recomputation at runtime. BatchNorm in the convolution module is folded into the depthwise conv weights at load time. Cross-attention K/V is pre-computed once per audio slice from the encoder output and then reused across decoder steps.

How this was made

1. Inspect the .nemo tarball: 1510 tensors total — encoder (1294), transf_decoder (214), log_softmax head (2), preprocessor (2). Skipped the auxiliary timestamps_asr_model_weights.ckpt which is the separate Parakeet CTC model used by NeMo Forced Aligner for segment-level timestamps.
2. Convert with models/convert-canary-to-gguf.py : remap NeMo state-dict keys ( transf_decoder._embedding.token_embedding → decoder.embed , first_sub_layer.query_net → sa_q , etc.) and write 1478 tensors as F16 (matmul) + F32 (norms / biases / mel filterbank). 1.97 GB GGUF.
3. C++ runtime in src/canary.{h,cpp} : mmap the GGUF, fold BN into the depthwise conv at load time, build the encoder graph (32-layer FastConformer with biases), build the decoder graph per step (with self-attention KV cache + pre-computed cross-K/V), greedy decode with task-token prompt, detokenise via SentencePiece.
4. Quantise with cohere-quantize : same llama.cpp-style quantiser used for the cohere and parakeet GGUFs in this fork.

Comparison with Parakeet TDT 0.6B v3

Attribution

• Original model: nvidia/canary-1b-v2 (CC-BY-4.0). NVIDIA NeMo team. See the Canary-1B-v2 & Parakeet-TDT-0.6B-v3 technical report .
• GGUF conversion + ggml runtime: CrispStrobe/cohere-whisper.cpp@parakeet . The decoder structure was cross-checked against NeMo's transformer_decoders.py and transformer_modules.py source.
• Encoder graph patterns: shared between cohere/parakeet/canary in the same fork, originally adapted for the cohere-whisper.cpp ggml branch .

Related

• C++ runtime: CrispStrobe/cohere-whisper.cpp@parakeet
• Sister model (ASR-only, smaller, with word timestamps): cstr/parakeet-tdt-0.6b-v3-GGUF
• Sister runtime (Cohere Transcribe, lowest English WER): cstr/cohere-transcribe-03-2026-GGUF
• ONNX INT4 (Cohere): cstr/cohere-transcribe-onnx-int4
• ONNX INT8 (Cohere): cstr/cohere-transcribe-onnx-int8

License

CC-BY-4.0, inherited from the base model. Use of these GGUF files must comply with the CC-BY-4.0 license including attribution.