diff --git a/icefall/shared/combine_lm.py b/icefall/shared/combine_lm.py
new file mode 100755
index 000000000..d7b85e1b4
--- /dev/null
+++ b/icefall/shared/combine_lm.py
@@ -0,0 +1,225 @@
+#!/usr/bin/env python3
+# flake8: noqa
+
+# Copyright    2023  Xiaomi Corp.        (author: Fangjun Kuang)
+#
+"""
+Given two LMs "A" and "B", this script modifies
+probabilities in A such that
+
+P_{A_{new}}(w_n|w_0,w_1,...,w_{n-1}) =  P_{A_{original}}(w_n|w_0,w_1,...,w_{n-1}) / P_{B}(w_n|w_0,w_1,...,w_{n-1})
+
+When it is formulated in log-space, it becomes
+
+\log P_{A_{new}}(w_n|w_0,w_1,...,w_{n-1}) =  \log P_{A_{original}}(w_n|w_0,w_1,...,w_{n-1}) - \log P_{B}(w_n|w_0,w_1,...,w_{n-1})
+
+Optionally, you can pass scales for LM "A" and LM "B", such that
+
+\log P_{A_{new}}(w_n|w0,w1,...,w_{n-1}) =  a_scale * \log P_{A_{original}}(w_n|w0,w1,...,w_{n-1}) - b_scale * \log P_{B}(w_n|w0,w1,...,w_{n-1})
+
+Usage:
+
+  python3 ./combine_lm.py \
+    --a 4-gram.arpa \
+    --b 2-gram.arpa \
+    --a-scale 1.0 \
+    --b-scale 1.0 \
+    --out new-4-gram.arpa
+
+It will generate a new arpa file `new-4-gram.arpa`
+"""
+import logging
+import re
+from typing import List
+
+try:
+    import kenlm
+except ImportError:
+    print("Please install kenlm first. You can use")
+    print()
+    print(" pip install https://github.com/kpu/kenlm/archive/master.zip")
+    print()
+    print("to install it")
+    import sys
+
+    sys.exit(-1)
+
+import argparse
+from pathlib import Path
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--a",
+        type=str,
+        required=True,
+        help="Path to the first LM. Its order is usually higher than that of LM b",
+    )
+
+    parser.add_argument(
+        "--b",
+        type=str,
+        required=True,
+        help="Path to the second LM. Its order is usually lower than that of LM a",
+    )
+
+    parser.add_argument(
+        "--a-scale",
+        type=float,
+        default=1.0,
+        help="Scale for the first LM.",
+    )
+
+    parser.add_argument(
+        "--b-scale",
+        type=float,
+        default=1.0,
+        help="Scale for the second LM.",
+    )
+
+    parser.add_argument(
+        "--out",
+        type=str,
+        required=True,
+        help="Path to save the generated LM.",
+    ),
+
+    return parser.parse_args()
+
+
+def check_args(args):
+    assert Path(args.a).is_file(), f"{args.a} does not exist"
+    assert Path(args.b).is_file(), f"{args.b} does not exist"
+
+
+def get_score(model: kenlm.Model, history: List[str], word: str):
+    """Compute \log_{10} p(word|history).
+
+    If history is [w0, w1, w2] and word is w3, the function returns
+    p(w3|w0,w1,w2)
+
+    Caution:
+      The returned score is in log10.
+
+    Args:
+      model:
+        The kenLM model.
+      history:
+        The history words.
+      word:
+        The current word.
+    Returns:
+      Return \log_{10} p(word|history).
+    """
+    order = model.order
+    history = history[-(order - 1) :]
+
+    in_state = kenlm.State()
+    out_state = kenlm.State()
+    model.NullContextWrite(in_state)
+
+    for w in history:
+        model.BaseScore(in_state, w, out_state)
+        in_state, out_state = out_state, in_state
+
+    return model.BaseScore(in_state, word, out_state)
+
+
+def _process_grams(
+    a: "_io.TextIOWrapper",
+    b: kenlm.Model,
+    a_scale: float,
+    b_scale: float,
+    order: int,
+    out: "_io.TextIOWrapper",
+):
+    """
+    Args:
+      a:
+       A file handle for the LM "A"
+      b:
+        LM B.
+      a_scale:
+        The scale for scores from LM A.
+      b_scale:
+        The scale for scores from LM B.
+      order: int
+        Current order of LM A.
+      out:
+        File handle for the output LM.
+    """
+    for line in a:
+        line = line.strip()
+        if not line:
+            print("", file=out)
+            break
+
+        s = line.strip().split()
+        assert len(s) > order, len(s)
+        assert len(s) >= order + 1, len(s)
+        assert len(s) <= order + 2, len(s)
+
+        log10_p_a = float(s[0])
+        history = s[1:order]
+        word = s[order]
+
+        log10_p_a_backoff = 0 if len(s) < order + 2 else float(s[-1])
+
+        log10_p_b = get_score(b, history, word)
+
+        # ensure that the resulting log10_p_a is not positive
+        log10_p_a = min(0, a_scale * log10_p_a - b_scale * log10_p_b)
+
+        log10_p_a_backoff *= a_scale
+
+        print(f"{log10_p_a:.7f}", end="\t", file=out)
+        if len(s) < order + 2:
+            print("\t".join(s[1:]), file=out)
+        else:
+            print("\t".join(s[1:-1]), end="\t", file=out)
+            print(f"{log10_p_a_backoff:.7f}", file=out)
+
+
+def process(args):
+    b = kenlm.LanguageModel(args.b)
+    logging.info(f"Order of {args.b}: {b.order}")
+    pattern = re.compile(r"\\(\d+)-grams:\n")
+    out = open(args.out, "w", encoding="utf-8")
+
+    a_scale = args.a_scale
+    b_scale = args.b_scale
+
+    with open(args.a, encoding="utf-8") as a:
+        for line in a:
+            print(line, end="", file=out)
+            m = pattern.match(line)
+            if m:
+                order = int(m.group(1))
+                _process_grams(
+                    a=a,
+                    b=b,
+                    a_scale=a_scale,
+                    b_scale=b_scale,
+                    order=order,
+                    out=out,
+                )
+    out.close()
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+    check_args(args)
+
+    process(args)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()