ThakiCloud/Voice2Policy-Text-Router-v0.1

Voice2Policy Text Router v0.1

🇰🇷 Korean card: MODEL_CARD.ko.md

A small character n-gram TF-IDF + logistic regression classifier that reads a Korean business voice command (an ASR transcript) and predicts its command frame — intent, action, and the kind of pending item — plus whether a booking time/date or phone number was mentioned. It is the routing step of ThakiCloud's Voice2Policy pipeline (ASR → this router → entity resolver → tool-schema check → policy engine).

It is a trained machine-learning model, but not a neural network, an LLM, a Transformer, or a speech model. It does not hear audio, does not resolve names, and does not execute anything.

Use

pip install numpy safetensors huggingface_hub

from huggingface_hub import snapshot_download
import sys
d = snapshot_download("ThakiCloud/Voice2Policy-Text-Router-v0.1")
sys.path.insert(0, d)
from predict import TextRouter

r = TextRouter(d)
print(r.predict("내일 오후 세 시에 스케일링 예약해 주세요"))
# {'intent': 'tool', 'action': 'book', 'kind': None, 'frame_score': 0.9991, ..., 'slot_mentioned': True, 'slot_score': ..., 'tool_hint': 'update_slots'}

Command line: python predict.py "콜드브루 라떼 두 잔 주세요" --json. No pickle and no scikit-learn are needed: weights are in model.safetensors, the n-gram vocabulary in vocab.json, settings and labels in config.json.

Output

field	meaning
intent	tool (business tool request), no_command (question, narration, refusal), delegate (hand off to staff: refund, complaint, discount, medical, asking for the owner)
action	book, order, confirm, decline (only when intent=tool)
kind	booking / order when the utterance says which pending item to confirm or cancel, else null
frame_score, frame_scores	softmax probability of the chosen frame and of all 10 frames
slot_mentioned, slot_score	a booking time/date or phone number was mentioned (sigmoid ≥ 0.5)
tool_hint	default tool family for the frame (update_slots, update_basket, confirm_*, decline_*, handoff)

From frame to tool. The final decision needs things this model does not see: the workspace's registered service/menu names, what is pending, and the tool's required arguments. In Voice2Policy that is done after the router (entity resolver, schema check, policy engine). Do not execute tool_hint directly.

Model

Features	scikit-learn TfidfVectorizer(analyzer="char_wb", ngram_range=(1,4), min_df=2, sublinear_tf=True, norm="l2") re-implemented in numpy — 8,716 character n-grams
Frame head	multinomial logistic regression, 10 classes (C = 10)
Slot head	binary logistic regression
Parameters	104,603 (coefficients + intercepts + IDF)
Files	model.safetensors 0.84 MB, vocab.json 0.15 MB
Inference	CPU, ~0.7 ms per utterance (measured inside the pipeline on one B200 host's CPU)
Training	10,682 sentences, 11 s on CPU

equivalence.json: the exported model reproduces the original scikit-learn model on 5,613 sentences — identical frame and slot decisions; probabilities differ only by the 4-decimal rounding of the output.

Training data

10,682 synthetic Korean business commands (booking / ordering / confirm / cancel / FAQ / narration / negation / hand-off) for 11 small-business types (clinic, cafe, restaurant, nail shop, …). The sentences were written by an LLM from typed command representations, and the labels come from deterministic rules on those representations, not from the LLM. No customer data, no recordings, and no personal information. The dataset is not released.

Evaluation (development comparison, single training run)

input	frame accuracy	intent+action	slot mentioned
held-out synthetic sentences (biz_test, 3,081)	0.998	0.998	0.999
ASR transcripts of the same test set read by TTS voices (18,879 clips; Qwen3-ASR-1.7B with registered names as context)	0.997	0.997	0.998
81 hand-written sentences in other wording	0.827	0.889	0.975

The held-out test shares the generator's style with training, so the first two rows are optimistic. On hand-written wording the router drops to 0.83. Typical misses: a refund request with no "staff" word is read as no_command; negated requests ("예약 취소하지 말아 주세요" — don't cancel) are read as the action; demonstratives alone ("그거 해 주세요" — do that) are guessed as a cancel. In Voice2Policy these are caught by the policy engine (negation on the tool verb, required action evidence for confirm/cancel, policy hand-off cues); with that layer the same 81 sentences give 76/81 correct decisions and no unsafe execution.

Limitations

• Korean only; vocabulary of small-business booking/ordering. Other domains need retraining.
• Synthetic training text only; real users' phrasing is not measured.
• A bag of character n-grams has no notion of scope: negation and reference need the downstream policy checks.
• Scores are not calibrated probabilities of being correct.

License

Apache-2.0 (weights, config and code in this repository).