johrell/S1-mini-sv

S1-mini by Superwhisper, Swedish by Babbla

A bilingual (Swedish + English) fine-tune of S1-mini by Superwhisper, the 0.6B text normalizer for speech-to-text transcripts, trained by Babbla to clean Swedish dictations the way it cleans English ones: fillers and false starts removed, self-corrections resolved, punctuation and capitalization applied, and spoken numbers, dates, clock times, amounts, phone numbers and addresses written out in Swedish written convention. English behaviour is preserved by replaying the original model's own outputs during training.

This repository is a derivative work of superwhisper/s1-mini (commit 88f6b15896c73bbb13a3b596e0afe8ea0d5150b4) and keeps its name, "S1-mini" by "Superwhisper", as the licence's naming clause requires. The Swedish capability is Babbla's addition. 8-bit MLX weights (group size 64), for Apple Silicon: measured, the 4-bit build of this fine-tune lost most of what it learned, 8-bit keeps it whole.

Input format

Identical to the original — the system prompt and control line are part of the trained format and must be sent verbatim:

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.

[Styling: semi-formal] [Structure: prose|lists] [Context: general|email]
<raw transcript>

Apply Qwen3's chat template with enable_thinking=False and decode greedily (generation_config.json ships do_sample: false). The model was fine-tuned under Styling: semi-formal only; other styling values were not part of the Swedish training data and may garble Swedish output.

Quickstart (mlx-lm)

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("johrell/S1-mini-sv")
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, structure="prose", context="general"):
    messages = [
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": f"[Styling: semi-formal] [Structure: {structure}] [Context: {context}]\n{transcript}"},
    ]
    prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, enable_thinking=False)
    return generate(model, tokenizer, prompt, max_tokens=512, sampler=make_sampler(temp=0.0))

print(normalize("Fakturan hamnade på tjugotretusen fyrahundrafemtio kronor och ska betalas senast den femtonde mars."))
# Fakturan hamnade på 23 450 kr och ska betalas senast den 15 mars.
print(normalize("Eh, vi behöver köpa mjölk, smör, ost och bröd.", structure="lists"))
# Vi behöver köpa:
# - Mjölk
# - Smör
# - Ost
# - Bröd

Swedish written convention

Svenska skrivregler with one exception: clock times take a colon (kl. 14:30), never a period, because 14.30 reads as a sentence end to downstream tools. Small counts in prose stay words (1–12), 13 and up are digits; anything with a unit, currency, percent, date or time is digits (23 450 kr, 1 250,50 kr, 20 %, 15 mars); halv tre becomes 14:30 when the half of day is clear and 2:30 otherwise; spelled addresses become addresses (anna punkt svensson snabel-a example punkt se → anna.svensson@example.se); self-corrections resolve to the final value. Lists and email layout copy the original model's English layout exactly.

Evaluation

Measured on Babbla's Swedish golden set (real KB-Whisper/Parakeet dictations plus authored cases, hand-corrected) with the 8-bit weights in this repository, greedy decoding:

tag	n	exact match	edit distance
all	124	89.5 %	0.005
fillers	19	84.2 %	0.015
punctuation	21	66.7 %	0.011
question	27	85.2 %	0.004
itn	32	93.8 %	0.002
list	16	87.5 %	0.013
self-correction	15	100.0 %	0.000
mixed-language	21	76.2 %	0.006
email	17	94.1 %	0.002

English agreement with the original S1-mini on 830 held-out pairs: 78.1 % exact.

Exact match is strict: a comma placed differently from the hand-corrected target counts as a miss, which is most of what the punctuation and mixed-language rows lose. For scale, the original S1-mini's own 8-bit and 4-bit builds agree 82 % with its bf16 greedy outputs on the same English pairs, so the 78 % above is the fine-tune's cost on English against that quantization ceiling, not against 100 %.

Training data

Generated backwards from clean Swedish text by rule, no audio involved: sentences from Riksdagen protocol text (Swedish open data) and Swedish Wikipedia (CC BY-SA 4.0), rendered into their spoken form (numbers, dates, times, amounts, addresses as words; fillers, false starts and self-corrections injected) with input habits matched to what KB-Whisper and Parakeet actually emit; about 2 000 spoken-style Swedish scripts written by Claude across Babbla's app categories; rule-built lists, step sequences, emails and command lines; and an English replay layer whose inputs come from Simple English Wikipedia plus the author's own English dictations and whose targets are the original S1-mini's greedy outputs. No Swedish user dictations were trained on: every real Swedish dictation the author had went into the held-out evaluation set instead.

Limitations

• Swedish and English only. Other languages were not trained and are handled no better than by the original.
• Only Styling: semi-formal was trained for Swedish; the other three styling values of the original are untested on Swedish input.
• No translation, no personal dictionary, no register for terminals or code editors: it normalizes, and a downstream app must own those.
• The retention behaviour of a text normalizer applies: it deletes fillers and false starts and may drop a hedge word it takes for one.
• English on bare, lowercase, unpunctuated input is a little weaker than the original: "so um i need to like send the the report by uh friday no wait make that thursday" comes back as "So i need to send the report by Thursday." (lowercase "i"; the original capitalizes it). Recognizer-style input, capitalized and punctuated, is what it was trained on.
• A 170-word single-sentence-free run-on is the longest shape it was trained on; keep inputs under roughly 1 000 tokens.

Licence

Apache 2.0 plus the naming clause of the original: see LICENSE and NOTICE. Any use or distribution, modified or not, must continue to identify the model as "S1-mini" by "Superwhisper", with that capitalization.