[](https://superwhisper.com)
[](https://discord.gg/tF98XvJNvB)
[](https://huggingface.co/superwhisper/s1-mini/tree/v1)
A 0.6B-parameter text normalizer for speech-to-text output. It takes a raw ASR
transcript and rewrites it as clean written text: fillers removed, false starts
and self-corrections resolved to the value the speaker landed on, punctuation
and capitalization applied, and spoken numbers, dates, times, currency and
email addresses rendered in written form.
On a held-out set of 7,519 English cases it reaches 94.8% token accuracy, and
the quantized build is a 462 MiB file that runs comfortably on a laptop CPU.
This is release v1, and it covers English only. S1-mini is not a chat model
and will not follow general instructions; it does one job, and you steer it
with a control line at the top of the input.
Fine-tuned from [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B). If
you want to run it in llama.cpp, Ollama, LM Studio, or anything else built on
llama.cpp, grab the GGUF builds from
[superwhisper/s1-mini-GGUF](https://huggingface.co/superwhisper/s1-mini-GGUF).
You can use it in your own dictation app too, just check the license first.
Releases are tagged, so you can pin one:
`from_pretrained("superwhisper/s1-mini", revision="v1")`.
## Model overview
| | |
|---|---|
| Type | Causal language model, fine-tuned for a single transformation task |
| Base model | [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) |
| Parameters | 596M total (0.44B non-embedding), embeddings tied |
| Layers | 28 |
| Attention heads | 16 for Q, 8 for KV (GQA) |
| Precision | BF16 |
| Recommended input | Up to ~1,000 tokens; chunk longer transcripts |
| Language | English |
| License | Apache 2.0 + naming clause ([LICENSE](LICENSE)) |
> [!NOTE]
> The Hub sidebar reports 0.8B parameters for this repo. `config.json` sets
> `tie_word_embeddings`, but `model.safetensors` still stores `lm_head.weight`
> as a materialized copy of the input embedding, so the 155.6M-parameter
> embedding is counted twice: 751.6M tensor elements against 596.0M unique
> parameters. The layout is inherited from `Qwen/Qwen3-0.6B`, which reports 0.8B
> on the Hub for the same reason. The table above counts unique parameters.
**Input.** The model expects the system prompt, then a control line, a
newline, and one raw ASR transcript, which will usually arrive lowercase and
unpunctuated. That is the shape it was trained on.
**Output.** It returns the cleaned transcript as plain text and nothing else,
with no preamble and no explanation. Under `Structure: lists` the output may
contain Markdown bullets, and under `Context: email` it may contain blank
lines separating a greeting, body and sign-off. When the input is nothing but
filler or noise, the correct output is an empty string, and that is what you
get.
## Quickstart
Qwen3 support landed in `transformers` 4.51.0; with anything older you will
get `KeyError: 'qwen3'`.
```bash
pip install "transformers>=4.51.0" torch
```
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL = "superwhisper/s1-mini"
tok = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForCausalLM.from_pretrained(MODEL, torch_dtype="auto")
# Required. Use this exact system prompt.
SYSTEM = (
"You are a text normalizer for speech-to-text transcripts. The input begins "
"with a control line specifying the styling, structure, and context settings; "
"clean the transcript to match those settings and output only the cleaned text."
)
def normalize(transcript, styling="semi-formal", structure="prose", context="general"):
control = f"[Styling: {styling}] [Structure: {structure}] [Context: {context}]"
messages = [
{"role": "system", "content": SYSTEM},
{"role": "user", "content": f"{control}\n{transcript}"},
]
text = tok.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=False, # required, see below
)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
return tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
raw = "so um i need to like send the the report by uh friday no wait make that thursday"
print(normalize(raw))
# I need to send the report by Thursday.
```
`torch_dtype="auto"` picks up the BF16 in `config.json`. At 0.6B this runs
comfortably on CPU; add `device_map="auto"` to place it on a GPU automatically
(requires `accelerate`).
## The control line
Every input starts with a control line, then a newline, then the transcript:
```
[Styling: