Speech recognition models

Pick a model that understands you.

Parroting is not tied to any one speech recognition model. During setup you pick one of the open models below to download; later you can switch any time from Manage models (Talk → Speech model & language). Whichever you choose, recognition runs on your Mac and your voice is never uploaded.

All run on your MacSwitch any timeYour dictionary works with every model

Not sure which one?

Three starting points, depending on your language and Mac. No speech model is marked “Recommended” for English users; pick the one that suits you.

Fast and smallSenseVoice Small

About 250 MB and the fastest multilingual model here. Runs smoothly on older Macs too.

Most accurate ChineseQwen3-ASR 1.7B

The fewest Chinese errors on this list, at about 1.5 GB and slower. Qwen3-ASR 0.6B is a middle ground.

Taiwan Mandarin with English mixed inBreeze-ASR-25

Tuned by MediaTek Research for Taiwan Mandarin and code-switching. The slowest, but it knows Taiwanese usage better.

Comparison

The error rates below are for Chinese, not English. They come from the public transcribe.cpp benchmarks on the FLEURS Chinese test set; speed was measured on an M4 Max GPU. Your Mac, voice and language will change the results.

Model	Developer	Languages	Download	Chinese error ratelower is better	Speedhigher is faster	License	Source
SenseVoice Small	FunAudioLLM (Alibaba)	Chinese, Cantonese, English, Japanese, Korean	253 MB	10.1%	267×	FunASR Model License	Original .gguf file
Qwen3-ASR 0.6B	Qwen (Alibaba)	30, incl. Chinese, Cantonese, English	850 MB	7.6%	69×	Apache 2.0	Original .gguf file
Fun-ASR Nano	FunAudioLLM (Alibaba)	Chinese, English, Japanese	891 MB	8.6%	78×	Apache 2.0	Original .gguf file
Breeze-ASR-25	MediaTek Research	Chinese (Taiwan), English	1.2 GB	8.1%	22×	Apache 2.0	Original .gguf file
Qwen3-ASR 1.7B	Qwen (Alibaba)	30, incl. Chinese, Cantonese, English	1.5 GB	7.1%	41×	Apache 2.0	Original .gguf file
Fun-ASR Nano Multilingual	FunAudioLLM (Alibaba)	31, incl. Chinese, Cantonese, English	891 MB	7.3%	70×	Apache 2.0	Original .gguf file
Whisper Large v3 Turbo	OpenAI	About 100	886 MB	8.5%	44×	MIT	Original .gguf file
Moonshine Tiny (Chinese)	Moonshine AI	Chinese	35 MB	13.9%	97×	Moonshine Community License	Original .gguf file
Moonshine Base (Chinese)	Moonshine AI	Chinese	77 MB	17.8%	67×	Moonshine Community License	Original .gguf file
SenseVoice	FunAudioLLM (Alibaba)	Chinese, Cantonese, English, Japanese, Korean	160 MB	—	—	FunASR Model License	Original Download file
Breeze ASR	MediaTek Research	Chinese (Taiwan), English	1.1 GB	—	—	Apache 2.0	Original Download file

The last two rows, “SenseVoice” and “Breeze ASR”, are older-format builds downloaded straight from Handy’s mirror; their model, authors and license are the same as SenseVoice Small and Breeze-ASR-25, and there are no public benchmark figures for them.

What each model is good at

SenseVoice SmallFunAudioLLM (Alibaba) · 253 MB

Small and the fastest; it detects whether you are speaking Chinese, Cantonese, English, Japanese or Korean. A few more Chinese errors than the larger models.

Qwen3-ASR 0.6BQwen (Alibaba) · 850 MB

30 languages, strong on Chinese and Cantonese, a good balance of accuracy and speed.

Fun-ASR NanoFunAudioLLM (Alibaba) · 891 MB

A newer FunAudioLLM model for Chinese, English and Japanese; Chinese accuracy close to Qwen3-ASR 0.6B.

Breeze-ASR-25MediaTek Research · 1.2 GB

Built on Whisper and tuned for Taiwan Mandarin and code-switching. The slowest on this list.

Qwen3-ASR 1.7BQwen (Alibaba) · 1.5 GB

The larger Qwen3-ASR; the most accurate here for Chinese and Cantonese, and it uses more memory.

Fun-ASR Nano MultilingualFunAudioLLM (Alibaba) · 891 MB

The multilingual Fun-ASR Nano: 31 languages, very accurate for Chinese and Japanese.

Whisper Large v3 TurboOpenAI · 886 MB

OpenAI’s multilingual model with the most languages; weaker on Cantonese.

Moonshine Tiny (Chinese)Moonshine AI · 35 MB

Tiny, about 35 MB, Chinese only. Noticeably less accurate; good when space is tight or just to try things out.

Moonshine Base (Chinese)Moonshine AI · 77 MB

About 77 MB, Chinese only. In this benchmark it is not more accurate than Tiny.

SenseVoiceFunAudioLLM (Alibaba) · 160 MB

An older-format build of SenseVoice Small (ONNX int8, runs on the CPU); same model and license. No public benchmark figures.

Breeze ASRMediaTek Research · 1.1 GB

An older-format build of Breeze-ASR-25 (ggml q5_k); same model and license. No public benchmark figures.

Use your own model

Models outside this list work too: any .gguf file of a kind Parroting supports (such as SenseVoice, Qwen3-ASR, Fun-ASR, Whisper or Moonshine). On the Talk page, open Speech model & language, click Manage models, then Import model… next to Downloaded Models. Parroting test-recognizes a sentence first and only switches once it gets the words; you can switch back any time. To train a model with your own recordings, see Learns from you.

How to import a model →Train a model with your own recordings →

Licenses

You download models right in the app, no files to hunt for: pick one during setup, and get others later from Manage models (Talk → Speech model & language). In the Source column, “Original” is the developer’s open-source page and “.gguf file” is the converted version the app downloads, if you want the details.

Each model has its own license, set by its developer; click a license name in the table to read it. Read it before you use or fine-tune a model; anything you train from these models still follows the original license.

You download these models from third-party hosts (Hugging Face and Handy’s mirror, blob.handy.computer) and use them under their own licenses; Parroting does not distribute or modify them and does not represent their developers. The .gguf conversions are provided by the Handy project (handy-computer).

A model you import yourself is used under the license of its source.

Guide: speech recognition models →Learns from you: train with your own voice →

Don’t have Parroting yet?

Download the free Mac app and pick a model during setup.

Download for Mac ↓

Free · Apple silicon · macOS 11 or later · v0.4.13 · about 50 MB