ThunderTalk ASR 0.6B

ThunderTalk's own speech-recognition model for dictation in Chinese, English and mixed Chinese–English. It is the default dictation model of ThunderTalk, a free voice input app for macOS, on Apple Silicon Macs with 16 GB or more.

Fine-tuned from Qwen/Qwen3-ASR-0.6B, then quantized: the text decoder is 8-bit and the audio encoder stays float16 (1.2 GB). It runs on the Mac's GPU (MLX).

Versions

Revision In ThunderTalk What changed
v2 (branch) 1.15.0 and later Trained further on names people dictate now (models, agent tools, developer, finance and academic terms) and on punctuation
v1 (main) 1.14.3 First 8-bit release

ThunderTalk pins the revision it was measured with and updates the model in the background when a new app version pins a newer one.

Results

Token error rate (CJK characters and Latin words; case, punctuation, numerals and fillers normalised) on ThunderTalk's standard test set: 44 recordings with human reference transcripts, decoded through the app's dictation path on an M3 Max. Lower is better.

Qwen3-ASR-0.6B v1 v2
All 8.7% 8.0% 7.7%
Short Chinese 6.9% 6.3% 6.3%
Short English 20.2% 17.2% 16.2%
Short mixed Chinese–English 16.1% 14.7% 13.3%
Long Chinese (1.7–4.4 min) 5.6% 4.4% 4.3%
Long mixed (1.1–3.1 min) 11.4% 10.8% 11.3%
Long English (1.6–2.9 min) 8.0% 8.5% 5.9%
Time per output token 36 ms 12 ms 12 ms

Names: 191 sentences that say 724 names, each read by two synthetic voices that were not used in training; none of the sentences was trained on. Share of names written correctly, higher is better.

v1 v2
All names 67.7% 80.4%
Developer and AI terms (Claude Code, Supabase, MLX, PyTorch…) 70.5% 86.4%
Agent tooling, finance, academic terms 80.0% 91.0%
Names from October–November 2026 (Gemini 4, Kimi K3.1, Lightpanda…) 51.4% 61.3%
  • Punctuation: on 40 real dictations v2 writes 306 full stops and question marks where v1 wrote 202 and joined the rest with commas.
  • Real speech: of 48 places in those dictations where v1 and the trained models differ and it is clear what was said, v1 has 18 right and v2 33.
  • v2 is the average of two training rounds merged at 0.7 strength. At full strength the rounds write more names (90% on the first two groups) but begin to hear English words in plain Chinese; 0.7 keeps most of the names without that.
  • A name the model has never seen is still often wrong (51–57% on the newest group before training on it). In ThunderTalk, hotwords and "Names on screen" cover those: 65 untrained names added as hotwords go from 54.7% to 69.3%, with no other word changed.

About 3× faster than the float16 base model: dictation decodes one token after another, and 8-bit weights halve the memory read for each one. The short test clips come from the test splits of public corpora similar to the training data, so expect a smaller gain on speech that differs from them.

Licence

Free for personal, educational, research and non-commercial use, for non-profits, and inside the ThunderTalk app for companies with fewer than 250 employees and under US$10 million in annual revenue. Credit ThunderTalk when you share the model or something built with it. Larger companies using it at work, building it into another commercial product or service, offering it as a hosted service or API, or redistributing it commercially need a written licence: song@entropinia.com. Full terms: LICENSE.

Based on Qwen3-ASR-0.6B by the Qwen team, Alibaba Cloud (Apache License 2.0, see LICENSE-Qwen). Fine-tuning data includes synthetic speech of ThunderTalk's own sentences, ASCEND (CC BY-SA 4.0) and FLEURS (CC BY 4.0).

Downloads last month
41
Safetensors
Model size
0.4B params
Tensor type
F16
·
U32
·
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Oulasong/ThunderTalk-ASR-0.6B

Finetuned
(103)
this model