diff --git a/requirements.txt b/requirements.txt
index f1afabb1..ba83326e 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,5 +1,6 @@
 black==22.3.0
 h5py
+librosa
 matplotlib
 numpy
 soundfile
diff --git a/returnn/hdf.py b/returnn/hdf.py
index e75b5736..8a4bbd63 100644
--- a/returnn/hdf.py
+++ b/returnn/hdf.py
@@ -4,6 +4,7 @@
 from enum import Enum, auto
 import glob
 import math
+import librosa
 import numpy as np
 import os
 import shutil
@@ -230,6 +231,7 @@ class RoundingScheme(Enum):
         "multi_channel_strategy": BaseStrategy(),
         "rounding": RoundingScheme.start_and_duration,
         "round_factor": 1,
+        "target_sampling_rate": None,
     }
 
     def __init__(
@@ -241,6 +243,7 @@ def __init__(
         returnn_root: Optional[tk.Path] = None,
         rounding: RoundingScheme = RoundingScheme.start_and_duration,
         round_factor: int = 1,
+        target_sampling_rate: Optional[int] = None,
     ):
         """
 
@@ -256,6 +259,7 @@ def __init__(
             start_and_duration will round down the start time and the duration of the segment
             rasr_compatible will round up the start time and round down the end time
         :param round_factor: do the rounding based on a sampling rate that is scaled down by this factor
+        :param target_sampling_rate: desired sampling rate for the HDF, data will be resampled to this rate if needed
         """
         self.set_vis_name("Dump audio to HDF")
         assert output_dtype in ["float64", "float32", "int32", "int16"]
@@ -267,6 +271,7 @@ def __init__(
         self.returnn_root = returnn_root
         self.rounding = rounding
         self.round_factor = round_factor
+        self.target_sampling_rate = target_sampling_rate
 
         self.out_hdf = self.output_path("audio.hdf")
 
@@ -284,7 +289,7 @@ def run(self):
 
         if self.segment_file:
             with uopen(self.segment_file, "rt") as f:
-                segments_whitelist = set(l.strip() for l in f.readlines() if len(l.strip()) > 0)
+                segments_whitelist = {line.strip() for line in f.readlines() if len(line.strip()) > 0}
         else:
             segments_whitelist = None
 
@@ -295,35 +300,46 @@ def run(self):
             audio = sf.SoundFile(audio_file)
 
             for segment in recording.segments:
-                if (not segments_whitelist) or (segment.fullname() in segments_whitelist):
-                    if self.rounding == self.RoundingScheme.start_and_duration:
-                        start = int(segment.start * audio.samplerate / self.round_factor) * self.round_factor
-                        duration = (
-                            int((segment.end - segment.start) * audio.samplerate / self.round_factor)
-                            * self.round_factor
-                        )
-                    elif self.rounding == self.RoundingScheme.rasr_compatible:
-                        start = math.ceil(segment.start * audio.samplerate / self.round_factor) * self.round_factor
-                        duration = (
-                            math.floor(segment.end * audio.samplerate / self.round_factor) * self.round_factor - start
-                        )
-                    else:
-                        raise NotImplementedError(f"RoundingScheme {self.rounding} not implemented.")
-                    audio.seek(start)
-                    data = audio.read(
-                        duration,
-                        always_2d=True,
-                        dtype=self.output_dtype,
+                if (segments_whitelist is not None) and (segment.fullname() not in segments_whitelist):
+                    continue
+
+                # determine correct start and duration values
+                if self.rounding == self.RoundingScheme.start_and_duration:
+                    start = int(segment.start * audio.samplerate / self.round_factor) * self.round_factor
+                    duration = (
+                        int((segment.end - segment.start) * audio.samplerate / self.round_factor) * self.round_factor
                     )
-                    if isinstance(self.multi_channel_strategy, self.PickNth):
-                        data = data[:, self.multi_channel_strategy.channel]
-                    else:
-                        assert data.shape[-1] == 1, "Audio has more than one channel, choose a multi_channel_strategy"
-                    out_hdf.insert_batch(
-                        inputs=data.reshape(1, -1, 1),
-                        seq_len=[data.shape[0]],
-                        seq_tag=[segment.fullname()],
+                elif self.rounding == self.RoundingScheme.rasr_compatible:
+                    start = math.ceil(segment.start * audio.samplerate / self.round_factor) * self.round_factor
+                    duration = (
+                        math.floor(segment.end * audio.samplerate / self.round_factor) * self.round_factor - start
                     )
+                else:
+                    raise NotImplementedError(f"RoundingScheme {self.rounding} not implemented.")
+
+                # read audio data
+                audio.seek(start)
+                data = audio.read(duration, always_2d=True, dtype=self.output_dtype)
+                if isinstance(self.multi_channel_strategy, self.PickNth):
+                    data = data[:, self.multi_channel_strategy.channel]
+                else:
+                    assert data.shape[-1] == 1, "Audio has more than one channel, choose a multi_channel_strategy"
+
+                # resample if necessary
+                if (sr := self.target_sampling_rate) is not None and sr != audio.samplerate:
+                    data = librosa.resample(
+                        y=data.astype(float),
+                        orig_sr=audio.samplerate,
+                        target_sr=sr,
+                        axis=0,
+                    ).astype(self.output_dtype)
+
+                # add audio to hdf
+                out_hdf.insert_batch(
+                    inputs=data.reshape(1, -1, 1),
+                    seq_len=[data.shape[0]],
+                    seq_tag=[segment.fullname()],
+                )
 
             audio.close()