#!/usr/bin/env python3 """Fleet-Clean: speech -> Fleet-2 transcript -> Laya word tagger -> clean text. python fleet_clean.py clip.wav # print the clean transcript python fleet_clean.py --json clip.wav other.wav # raw + clean + per-word tags as JSON lines python fleet_clean.py --text "so I I think, you know, we should go" Speech runs in Cactus' native engine on whistle.cact (pip install cactus-compute); cleanup runs the tagger in PyTorch (pip install torch "transformers>=5" safetensors). Both models sit next to this file. The tagger only deletes words and changes punctuation / first-letter case: it never writes a word of its own. """ import argparse import json import sys from pathlib import Path HERE = Path(__file__).resolve().parent sys.path.insert(0, str(HERE)) from fleet_clean_words import C_KEEP, PUNCT_MODES, apply_tags, resolve, trailing_class # noqa: E402 class Tagger: def __init__(self, path=HERE / "tagger", device=None): import torch from safetensors.torch import load_file from transformers import AutoTokenizer, ModernBertModel self.torch = torch self.device = device or ("mps" if torch.backends.mps.is_available() else "cpu") self.encoder = ModernBertModel.from_pretrained(path).to(self.device).eval() head = load_file(str(Path(path) / "head.safetensors")) self.w, self.b = head["weight"].to(self.device), head["bias"].to(self.device) self.tok = AutoTokenizer.from_pretrained(path) import json labels = json.loads((Path(path) / "labels.json").read_text()) self.threshold = labels.get("delete_threshold", 0.5) self.punct_mode = labels.get("punct_mode", "boundary") def tag(self, words, delete_threshold=None): delete_threshold = self.threshold if delete_threshold is None else delete_threshold torch = self.torch if not words: return [], [], [] enc = self.tok(words, is_split_into_words=True, truncation=True, max_length=512) first = [None] * len(words) for t, w in enumerate(enc.word_ids()): if w is not None and first[w] is None: first[w] = t x = torch.tensor([enc["input_ids"]], device=self.device) with torch.no_grad(): h = self.encoder(input_ids=x, attention_mask=torch.ones_like(x)).last_hidden_state[0].float() logits = (h @ self.w.T + self.b).cpu() p_del = logits[:, :2].softmax(-1)[:, 1] d, p, c = [], [], [] for w, t in zip(words, first): if t is None: # past the 512-token window: keep the word as written d.append(0), p.append(trailing_class(w)), c.append(C_KEEP) else: d.append(int(p_del[t] >= delete_threshold)) p.append(int(logits[t, 2:7].argmax())) c.append(int(logits[t, 7:10].argmax())) return d, p, c def clean(self, text, delete_threshold=None, punct_mode=None): words = text.split() d, p, c = self.tag(words, delete_threshold) P, C = resolve(words, d, p, c, punct_mode or self.punct_mode) return apply_tags(words, d, P, C), {"words": words, "delete": d, "punct": P, "case": C} def transcribe(path, weights=HERE / "whistle.cact", language=None): import cactus res = cactus.transcribe(str(path), language=language, weights=str(weights)) return res["text"].strip(), res.get("language") def main(argv=None): ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) ap.add_argument("audio", nargs="*", help="WAV files (16 kHz mono; other rates need cactus-compute[mic])") ap.add_argument("--text", help="clean this text instead of transcribing audio") ap.add_argument("--language", default=None, help="en de fr es it nl pl ko (default: detect)") ap.add_argument("--threshold", type=float, default=None, help="delete probability threshold (default: calibrated)") ap.add_argument("--punct", choices=PUNCT_MODES, default=None, help="boundary (default): fix punctuation only where words were removed; full: re-punctuate everything") ap.add_argument("--device", default=None) ap.add_argument("--json", action="store_true") a = ap.parse_args(argv) if not a.text and not a.audio: ap.error("give a WAV file or --text") tagger = Tagger(device=a.device) jobs = [("text", a.text)] if a.text else [("audio", p) for p in a.audio] for kind, item in jobs: raw, lang = transcribe(item, language=a.language) if kind == "audio" else (item, a.language) clean, tags = tagger.clean(raw, a.threshold, a.punct) if a.json: print(json.dumps({"source": str(item) if kind == "audio" else None, "language": lang, "raw": raw, "clean": clean, **tags}, ensure_ascii=False)) else: print(clean) if __name__ == "__main__": main()