cswk/sayit-asr-gguf

SayIt ASR — GGUF weights

The local speech-recognition models shipped with SayIt, an open-source voice typing app for Windows: press a shortcut, speak, and AI-cleaned text lands at your cursor.

These seven files are byte-identical re-uploads of the handy-computer GGUF conversions, mirrored here so SayIt's model downloader does not depend on a third-party repository staying online. No requantization, no retraining, no modification of any kind.

Acknowledgements

All the hard work here belongs to Handy and @cjpais.

SayIt's entire local speech-recognition stack rests on two things they built and gave away: transcribe.cpp, the ggml-based C/C++ inference library (MIT), and the handy-computer model org, which converted these ASR models to GGUF and keeps a much larger catalog than the seven files here. Without those conversions there would be no local mode in SayIt at all — running SenseVoice, Fun-ASR, Qwen3-ASR, Parakeet and Nemotron on consumer hardware through one runtime is the difficult part, and it was already solved.

If you are looking for the canonical source of any file in this repo, go upstream:

This repo	Upstream conversion
parakeet-unified-en-0.6b-Q4_K_M.gguf	handy-computer/parakeet-unified-en-0.6b-gguf
SenseVoiceSmall-Q8_0.gguf	handy-computer/SenseVoiceSmall-gguf
nemotron-3.5-asr-streaming-0.6b-Q4_K_M.gguf	handy-computer/nemotron-3.5-asr-streaming-0.6b-gguf
Fun-ASR-Nano-2512-Q8_0.gguf	handy-computer/Fun-ASR-Nano-2512-gguf
Qwen3-ASR-0.6B-Q8_0.gguf	handy-computer/Qwen3-ASR-0.6B-gguf
Qwen3-ASR-1.7B-Q4_K_M.gguf	handy-computer/Qwen3-ASR-1.7B-gguf
Qwen3-ASR-1.7B-Q5_K_M.gguf	handy-computer/Qwen3-ASR-1.7B-gguf

Also with thanks to the labs that trained and released the original models: NVIDIA (Parakeet, Nemotron), Alibaba FunAudioLLM (SenseVoice, Fun-ASR Nano) and Alibaba Qwen (Qwen3-ASR).

Compatibility

These are transcribe.cpp GGUF files. They are not loadable by llama.cpp or whisper.cpp. The general.architecture tags (sensevoice, funasr_nano, parakeet, qwen3_asr) are specific to that runtime, so a same-named GGUF from another uploader will fail to load even though the download and checksum succeed.

Models

Ordered fast to slow.

File	Base model	Quant	Size	Languages	RAM in use
parakeet-unified-en-0.6b-Q4_K_M.gguf	nvidia/parakeet-unified-en-0.6b	Q4_K_M	477 MB	English only	~0.9 GB
SenseVoiceSmall-Q8_0.gguf	FunAudioLLM/SenseVoiceSmall	Q8_0	253 MB	zh, en, ja, ko, yue	~0.35 GB
nemotron-3.5-asr-streaming-0.6b-Q4_K_M.gguf	nvidia/nemotron-3.5-asr-streaming-0.6b	Q4_K_M	496 MB	32 locales	~1.05 GB
Fun-ASR-Nano-2512-Q8_0.gguf	FunAudioLLM/Fun-ASR-Nano-2512	Q8_0	891 MB	zh, en, ja	~1.4 GB
Qwen3-ASR-0.6B-Q8_0.gguf	Qwen/Qwen3-ASR-0.6B	Q8_0	850 MB	30+	~1.5 GB
Qwen3-ASR-1.7B-Q4_K_M.gguf	Qwen/Qwen3-ASR-1.7B	Q4_K_M	1.32 GB	30+	~2.3 GB
Qwen3-ASR-1.7B-Q5_K_M.gguf	Qwen/Qwen3-ASR-1.7B	Q5_K_M	1.52 GB	30+	~2.6 GB

RAM in use is a measured resident working set while the model is loaded, not the file size. Autoregressive models such as Qwen3-ASR add 700 MB to 1 GB on top of the weights for KV cache and compute buffers.

Which one to pick

• English only → parakeet-unified-en-0.6b. In English it is simultaneously the fastest and the most accurate file here (upstream LibriSpeech test-clean WER 1.62% at Q4_K_M), so it dominates every other entry. It returns an empty string on Chinese audio, which is why it cannot be a general-purpose default.
• Smallest footprint → SenseVoiceSmall. 253 MB on disk, ~350 MB resident.
• Chinese dictation, best value → Fun-ASR-Nano-2512. Beats Qwen3-ASR 0.6B on both speed and accuracy (upstream WER 1.79% vs 2.11%) at roughly half the decode cost of the 1.7B, but only covers zh / en / ja.
• Widest language coverage → nemotron-3.5-asr-streaming-0.6b, 32 locales including Spanish, French, German, Italian, Portuguese, Dutch, Russian, Arabic, Hindi, Turkish, Vietnamese and Ukrainian. Noticeably less accurate in English than parakeet (WER 3.28% vs 1.62%), so pick it for coverage, not for English.
• Highest accuracy across many languages → Qwen3-ASR-1.7B-Q5_K_M (upstream WER 1.65%). The Q4_K_M variant trades WER 1.65% → 1.81% for 188 MB less download and lower memory pressure.

All WER figures above are the values reported upstream for these quantizations, not our own measurements.

Usage notes

Three behaviours that are easy to misdiagnose as bugs:

1. Punctuation on SenseVoice and Fun-ASR Nano is behind the ITN flag. With itn=off the model takes the <|woitn|> branch and returns unpunctuated text.
2. Nemotron 3.5 only accepts region-qualified language codes (en-US, zh-CN). Passing a bare en returns unsupported language (status 10). Every other model here takes bare codes. SayIt resolves this in gguf_asr.rs::resolve_language (exact match → first locale of the same language → fall back to auto-detect).
3. Parakeet and Nemotron ignore the ITN/PnC options — both report unsupported. Casing and punctuation are inherent model behaviour. Parakeet normalizes numbers (2026, not spelled out); Nemotron keeps them in spoken form, which is a training choice rather than a defect.

With SayIt

Nothing to do. Open SayIt → Settings → Voice engine → Local mode, and pick a model. It downloads, verifies, and loads it, with Vulkan GPU acceleration when a GPU is available (AMD and Intel included, not just NVIDIA).

Direct download

huggingface-cli download cswk/sayit-asr-gguf SenseVoiceSmall-Q8_0.gguf --local-dir .

Checksums

SHA-256, for verifying a manual download:

File	SHA-256
parakeet-unified-en-0.6b-Q4_K_M.gguf	a8bf3de2b393bd14ead5a858c3748d5e3b07a20fdeabdd3b498fba4f463fa929
SenseVoiceSmall-Q8_0.gguf	6c759ee4c9748c9b3f7a5a60ca74f0f7e685fb9d45d1378fce7cfd62f59adf29
nemotron-3.5-asr-streaming-0.6b-Q4_K_M.gguf	41c99fa5fb6f3d35f68e79adc3e755eca2232a8d921178bd647b71194792b8fd
Fun-ASR-Nano-2512-Q8_0.gguf	681caef6df15a2c0e153b40ca7fe4087fdf65751fa5e6fe605d8a75dff969e61
Qwen3-ASR-0.6B-Q8_0.gguf	f081b2d5e23bd669d92cc331d722a8a0681943b8e6f34b48996fd5c319b5acd8
Qwen3-ASR-1.7B-Q4_K_M.gguf	b7afe3674f653fa84f712ed2440353c6e7cf7f93697fef76b05a26538b24844e
Qwen3-ASR-1.7B-Q5_K_M.gguf	034c557fe92ff8fcd9a9c041cbdaad347be0a86a58d3a348f63cf3f0180879d0

Licenses

This repository carries a mix of upstream licenses. Each file stays under the license of its original model — re-hosting does not change it. Check the one you plan to use:

File	License	Holder
parakeet-unified-en-0.6b-Q4_K_M.gguf	CC-BY-4.0 (attribution required)	NVIDIA
nemotron-3.5-asr-streaming-0.6b-Q4_K_M.gguf	NVIDIA model license (other)	NVIDIA
SenseVoiceSmall-Q8_0.gguf	see upstream model card (other)	Alibaba FunAudioLLM
Fun-ASR-Nano-2512-Q8_0.gguf	see upstream model card (other)	Alibaba FunAudioLLM
Qwen3-ASR-0.6B-Q8_0.gguf	Apache-2.0	Alibaba Qwen
Qwen3-ASR-1.7B-*.gguf	Apache-2.0	Alibaba Qwen

SayIt itself is AGPL-3.0. That license applies to the application, not to these weights.