move out of metrics.

Author: vincentqb

docs/source/index.rst

@@ -12,7 +12,6 @@ The :mod:`torchaudio` package consists of I/O, popular datasets and common audio
    functional
    transforms
    datasets
-   metrics
    models
    sox_effects
    compliance.kaldi

docs/source/metrics.rst

@@ -1,5 +1,5 @@
 from parameterized import parameterized
-from torchaudio.metrics import levenshtein_distance
+from torchaudio.functional import levenshtein_distance
 from torchaudio_unittest import common_utils
 
 

test/torchaudio_unittest/metrics.py

@@ -4,7 +4,6 @@
     compliance,
     datasets,
     kaldi_io,
-    metrics,
     utils,
     sox_effects,
     transforms

torchaudio/__init__.py

@@ -2,6 +2,7 @@
 
 import math
 from typing import Optional, Tuple
+from collections.abc import Sequence
 import warnings
 
 import torch
@@ -41,6 +42,7 @@
     'mask_along_axis_iid',
     'sliding_window_cmn',
     'vad',
+    'levenshtein_distance',
 ]
 
 
@@ -2330,3 +2332,42 @@ def vad(
     res = waveform[:, pos - samplesLen_ns + flushedLen_ns:]
     # unpack batch
     return res.view(shape[:-1] + res.shape[-1:])
+
+
+def levenshtein_distance(r: Sequence, h: Sequence) -> int:
+    """
+    Calculate the Levenshtein distance between two sequences.
+
+    The function computes an edit distance allowing deletion, insertion and substitution.
+    The result is an integer. Users may want to normalize by the length of the reference.
+
+    This can be used to compute the edit distance for instance between two strings,
+    or two list of words. Note that, if a string and a list of words is provided, the distance
+    will be computed between the "string" sequence, and the "list of words" sequence.
+
+    Args:
+        r (Sequence): the reference sequence to compare.
+        h (Sequence): the hypothesis sequence (e.g. the predicted sequence) to compare.
+    Returns:
+        int: The distance between the reference and the hypothesis.
+    """
+
+    # Initialisation
+    dold = list(range(len(h) + 1))
+    dnew = list(0 for _ in range(len(h) + 1))
+
+    # Computation
+    for i in range(1, len(r) + 1):
+        dnew[0] = i
+        for j in range(1, len(h) + 1):
+            if r[i - 1] == h[j - 1]:
+                dnew[j] = dold[j - 1]
+            else:
+                substitution = dold[j - 1] + 1
+                insertion = dnew[j - 1] + 1
+                deletion = dold[j] + 1
+                dnew[j] = min(substitution, insertion, deletion)
+
+        dnew, dold = dold, dnew
+
+    return dold[-1]