diff --git a/.gitignore b/.gitignore
index 063a0eaa..0cf77b56 100644
--- a/.gitignore
+++ b/.gitignore
@@ -9,3 +9,4 @@ venv/
 __init__.pyc
 coverage.xml
 docs/api/
+benchmarks/cache/
diff --git a/benchmarks/README.md b/benchmarks/README.md
index bebf562a..2b38d192 100644
--- a/benchmarks/README.md
+++ b/benchmarks/README.md
@@ -13,6 +13,14 @@ on the following machine:
 * Linux: Ubuntu 22.04
 * Python: 3.10
 
+Before running any benchmark,
+make sure to install missing requirements with:
+
+```bash
+$ cd benchmarks/
+$ pip install -r requirements.txt
+```
+
 
 ## audb.Dependencies methods
 
@@ -23,39 +31,52 @@ This benchmark provides insights
 how to best represent
 the dependency table internally.
 
-Results for a dependency table containing 1,000,000 files
-represented by `pandas.DataFrame`
-as of commit 4bbcc07.
-
-| Method                                         | Execution time |
-| ---------------------------------------------- | -------------- |
-| `Dependency.__call__()`                        |        0.000 s |
-| `Dependency.__contains__()`                    |        0.000 s |
-| `Dependency.__get_item__()`                    |        0.000 s |
-| `Dependency.__len__()`                         |        0.000 s |
-| `Dependency.__str__()`                         |        0.006 s |
-| `Dependency.archives`                          |        0.147 s |
-| `Dependency.attachments`                       |        0.045 s |
-| `Dependency.attachment_ids`                    |        0.045 s |
-| `Dependency.files`                             |        0.185 s |
-| `Dependency.media`                             |        0.264 s |
-| `Dependency.removed_media`                     |        0.250 s |
-| `Dependency.table_ids`                         |        0.053 s |
-| `Dependency.tables`                            |        0.046 s |
-| `Dependency.archive(1000 files)`               |        0.005 s |
-| `Dependency.bit_depth(1000 files)`             |        0.004 s |
-| `Dependency.channels(1000 files)`              |        0.004 s |
-| `Dependency.checksum(1000 files)`              |        0.004 s |
-| `Dependency.duration(1000 files)`              |        0.004 s | 
-| `Dependency.format(1000 files)`                |        0.004 s |
-| `Dependency.removed(1000 files)`               |        0.004 s |
-| `Dependency.sampling_rate(1000 files)`         |        0.004 s |
-| `Dependency.type(1000 files)`                  |        0.005 s |
-| `Dependency.version(1000 files)`               |        0.004 s |
-| `Dependency._add_attachment()`                 |        0.061 s |
-| `Dependency._add_media(1000 files)`            |        0.050 s |
-| `Dependency._add_meta()`                       |        0.124 s |
-| `Dependency._drop()`                           |        0.078 s |
-| `Dependency._remove()`                         |        0.068 s |
-| `Dependency._update_media()`                   |        0.073 s |
-| `Dependency._update_media_version(1000 files)` |        0.008 s |
+To run the benchmark execute:
+
+```bash
+$ python benchmark-dependencies-methods.py
+```
+
+Execution times in seconds
+of `audb.Dependencies` methods
+for a dependency table
+containing 1,000,000 files
+stored as a `pandas.DataFrame`
+using different dtype representations
+(storing string as `string`,
+storing string as `object`,
+using `pyarrow` dtypes)
+as of commit 91528e4.
+
+| method                                         |   string |   object |   pyarrow |
+|------------------------------------------------|----------|----------|-----------|
+| Dependencies.\_\_call__()                      |    0.000 |    0.000 |     0.000 |
+| Dependencies.\_\_contains__()                  |    0.000 |    0.000 |     0.000 |
+| Dependencies.\_\_get_item__()                  |    0.000 |    0.000 |     0.000 |
+| Dependencies.\_\_len__()                       |    0.000 |    0.000 |     0.000 |
+| Dependencies.\_\_str__()                       |    0.006 |    0.005 |     0.007 |
+| Dependencies.archives                          |    0.141 |    0.116 |     0.144 |
+| Dependencies.attachments                       |    0.029 |    0.018 |     0.017 |
+| Dependencies.attachment_ids                    |    0.029 |    0.018 |     0.017 |
+| Dependencies.files                             |    0.030 |    0.012 |     0.043 |
+| Dependencies.media                             |    0.127 |    0.072 |     0.086 |
+| Dependencies.removed_media                     |    0.117 |    0.069 |     0.081 |
+| Dependencies.table_ids                         |    0.037 |    0.026 |     0.023 |
+| Dependencies.tables                            |    0.028 |    0.017 |     0.017 |
+| Dependencies.archive(1000 files)               |    0.005 |    0.005 |     0.007 |
+| Dependencies.bit_depth(1000 files)             |    0.004 |    0.004 |     0.006 |
+| Dependencies.channels(1000 files)              |    0.004 |    0.004 |     0.006 |
+| Dependencies.checksum(1000 files)              |    0.004 |    0.004 |     0.006 |
+| Dependencies.duration(1000 files)              |    0.004 |    0.004 |     0.006 |
+| Dependencies.format(1000 files)                |    0.004 |    0.004 |     0.006 |
+| Dependencies.removed(1000 files)               |    0.004 |    0.004 |     0.006 |
+| Dependencies.sampling_rate(1000 files)         |    0.004 |    0.004 |     0.006 |
+| Dependencies.type(1000 files)                  |    0.004 |    0.004 |     0.006 |
+| Dependencies.version(1000 files)               |    0.004 |    0.004 |     0.006 |
+| Dependencies._add_attachment()                 |    0.055 |    0.056 |     0.207 |
+| Dependencies._add_media(1000 files)            |    0.049 |    0.050 |     0.060 |
+| Dependencies._add_meta()                       |    0.120 |    0.128 |     0.138 |
+| Dependencies._drop()                           |    0.075 |    0.075 |     0.117 |
+| Dependencies._remove()                         |    0.068 |    0.068 |     0.064 |
+| Dependencies._update_media()                   |    0.071 |    0.072 |     0.125 |
+| Dependencies._update_media_version(1000 files) |    0.008 |    0.008 |     0.017 |
diff --git a/benchmarks/benchmark-dependencies-methods.py b/benchmarks/benchmark-dependencies-methods.py
new file mode 100644
index 00000000..12c23da4
--- /dev/null
+++ b/benchmarks/benchmark-dependencies-methods.py
@@ -0,0 +1,371 @@
+import hashlib
+import os
+import pickle
+import random
+import string
+import time
+
+import pandas as pd
+import tabulate
+
+import audeer
+
+import audb
+
+
+random.seed(1)
+
+cache = audeer.mkdir("./cache")
+
+
+def astype(df, dtype):
+    """Convert to desired dataframe dtypes."""
+    if dtype == "object":
+        # Use `object` to represent strings
+        df["archive"] = df["archive"].astype("object")
+        df["bit_depth"] = df["bit_depth"].astype("int32")
+        df["channels"] = df["channels"].astype("int32")
+        df["checksum"] = df["checksum"].astype("object")
+        df["duration"] = df["duration"].astype("float64")
+        df["format"] = df["format"].astype("object")
+        df["removed"] = df["removed"].astype("int32")
+        df["sampling_rate"] = df["sampling_rate"].astype("int32")
+        df["type"] = df["type"].astype("int32")
+        df["version"] = df["version"].astype("object")
+        df.index = df.index.astype("object")
+        # Set dtypes in library
+        audb.core.define.DEPEND_FIELD_DTYPES = {
+            "archive": "object",
+            "bit_depth": "int32",
+            "channels": "int32",
+            "checksum": "object",
+            "duration": "float64",
+            "format": "object",
+            "removed": "int32",
+            "sampling_rate": "int32",
+            "type": "int32",
+            "version": "object",
+        }
+    elif dtype == "string":
+        # Use `string` to represent strings
+        df["archive"] = df["archive"].astype("string")
+        df["bit_depth"] = df["bit_depth"].astype("int32")
+        df["channels"] = df["channels"].astype("int32")
+        df["checksum"] = df["checksum"].astype("string")
+        df["duration"] = df["duration"].astype("float64")
+        df["format"] = df["format"].astype("string")
+        df["removed"] = df["removed"].astype("int32")
+        df["sampling_rate"] = df["sampling_rate"].astype("int32")
+        df["type"] = df["type"].astype("int32")
+        df["version"] = df["version"].astype("string")
+        df.index = df.index.astype("string")
+        # Set dtypes in library
+        audb.core.define.DEPEND_FIELD_DTYPES = {
+            "archive": "string",
+            "bit_depth": "int32",
+            "channels": "int32",
+            "checksum": "string",
+            "duration": "float64",
+            "format": "string",
+            "removed": "int32",
+            "sampling_rate": "int32",
+            "type": "int32",
+            "version": "string",
+        }
+    elif dtype == "pyarrow":
+        # Use `pyarrow` to represent all dtypes
+        df["archive"] = df["archive"].astype("string[pyarrow]")
+        df["bit_depth"] = df["bit_depth"].astype("int32[pyarrow]")
+        df["channels"] = df["channels"].astype("int32[pyarrow]")
+        df["checksum"] = df["checksum"].astype("string[pyarrow]")
+        df["duration"] = df["duration"].astype("float64[pyarrow]")
+        df["format"] = df["format"].astype("string[pyarrow]")
+        df["removed"] = df["removed"].astype("int32[pyarrow]")
+        df["sampling_rate"] = df["sampling_rate"].astype("int32[pyarrow]")
+        df["type"] = df["type"].astype("int32[pyarrow]")
+        df["version"] = df["version"].astype("string[pyarrow]")
+        df.index = df.index.astype("string[pyarrow]")
+        # Set dtypes in library
+        audb.core.define.DEPEND_FIELD_DTYPES = {
+            "archive": "string[pyarrow]",
+            "bit_depth": "int32[pyarrow]",
+            "channels": "int32[pyarrow]",
+            "checksum": "string[pyarrow]",
+            "duration": "float64[pyarrow]",
+            "format": "string[pyarrow]",
+            "removed": "int32[pyarrow]",
+            "sampling_rate": "int32[pyarrow]",
+            "type": "int32[pyarrow]",
+            "version": "string[pyarrow]",
+        }
+    return df
+
+
+# === Dependencies pandas.DataFrame ===
+data_cache = audeer.path(cache, "df.pkl")
+num_rows = 1000000
+if not os.path.exists(data_cache):
+    bit_depths = [0, 16, 24]
+    channels = [0, 1, 2]
+    formats = ["csv", "wav", "txt"]
+    sampling_rates = [0, 16000, 44100]
+    types = [0, 1, 2]
+    versions = ["1.0.0", "1.1.0"]
+    records = [
+        {
+            "file": f"file-{n}.wav",
+            "archive": f"archive-{n}",
+            "bit_depth": random.choices(bit_depths, weights=[0.1, 0.8, 0.1])[0],
+            "channels": random.choices(channels, weights=[0.1, 0.8, 0.1])[0],
+            "checksum": hashlib.md5(
+                pickle.dumps(random.choice(string.ascii_letters))
+            ).hexdigest(),
+            "duration": 10 * random.random(),
+            "format": random.choices(formats, weights=[0.1, 0.8, 0.1])[0],
+            "removed": random.choices([0, 1], weights=[0.1, 0.9])[0],
+            "sampling_rate": random.choices(sampling_rates, weights=[0.1, 0.8, 0.1])[0],
+            "type": random.choices(types, weights=[0.1, 0.8, 0.1])[0],
+            "version": random.choices(versions, weights=[0.2, 0.8])[0],
+        }
+        for n in range(num_rows)
+    ]
+    df = pd.DataFrame.from_records(records)
+    for column, dtype in zip(
+        audb.core.define.DEPEND_FIELD_NAMES.values(),
+        audb.core.define.DEPEND_FIELD_DTYPES.values(),
+    ):
+        df[column] = df[column].astype(dtype)
+    df.set_index("file", inplace=True)
+    df.index.name = None
+    df.index = df.index.astype("string")
+    df.to_pickle(data_cache)
+
+
+# ===== Benchmark audb.Dependencies =====
+deps = audb.Dependencies()
+deps.load(data_cache)
+file = "file-10.wav"
+n_files = 1000
+_files = deps._df.index[:n_files].tolist()
+dtypes = ["string", "object", "pyarrow"]
+results = pd.DataFrame(columns=dtypes)
+results.index.name = "method"
+
+for dtype in dtypes:
+    deps.load(data_cache)
+    deps._df = astype(deps._df, dtype)
+
+    # Check we have the expected dtypes
+    # in dependency table
+    # and library
+    if dtype == "pyarrow":
+        expected_dtype = "string[pyarrow]"
+    else:
+        expected_dtype = dtype
+    assert deps._df.archive.dtype == expected_dtype
+    assert audb.core.define.DEPEND_FIELD_DTYPES["archive"] == expected_dtype
+
+    method = "Dependencies.__call__()"
+    t0 = time.time()
+    deps()
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    # Access the index one time.
+    # Further calls will be faster
+    file in deps
+
+    method = "Dependencies.__contains__()"
+    t0 = time.time()
+    file in deps
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.__get_item__()"
+    t0 = time.time()
+    deps[file]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.__len__()"
+    t0 = time.time()
+    len(deps)
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.__str__()"
+    t0 = time.time()
+    str(deps)
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.archives"
+    t0 = time.time()
+    deps.archives
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.attachments"
+    t0 = time.time()
+    deps.attachments
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.attachment_ids"
+    t0 = time.time()
+    deps.attachment_ids
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.files"
+    t0 = time.time()
+    deps.files
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.media"
+    t0 = time.time()
+    deps.media
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.removed_media"
+    t0 = time.time()
+    deps.removed_media
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.table_ids"
+    t0 = time.time()
+    deps.table_ids
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies.tables"
+    t0 = time.time()
+    deps.tables
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.archive({n_files} files)"
+    t0 = time.time()
+    [deps.archive(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.bit_depth({n_files} files)"
+    t0 = time.time()
+    [deps.bit_depth(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.channels({n_files} files)"
+    t0 = time.time()
+    [deps.channels(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.checksum({n_files} files)"
+    t0 = time.time()
+    [deps.checksum(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.duration({n_files} files)"
+    t0 = time.time()
+    [deps.duration(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.format({n_files} files)"
+    t0 = time.time()
+    [deps.format(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.removed({n_files} files)"
+    t0 = time.time()
+    [deps.removed(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.sampling_rate({n_files} files)"
+    t0 = time.time()
+    [deps.sampling_rate(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.type({n_files} files)"
+    t0 = time.time()
+    [deps.type(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies.version({n_files} files)"
+    t0 = time.time()
+    [deps.version(file) for file in _files]
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    # -------------------------------------------------------------------------
+    method = "Dependencies._add_attachment()"
+    t0 = time.time()
+    deps._add_attachment("attachment.txt", "1.0.0", "archive", "checksum")
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies._add_media({n_files} files)"
+    values = [
+        (
+            f"file-new-{n}.wav",  # file
+            f"archive-new-{n}",  # archive
+            16,  # bit_depth
+            1,  # channels
+            f"checksum-{n}",  # checksum
+            0.4,  # duration
+            "wav",  # format
+            0,  # removed
+            16000,  # sampling_rate
+            1,  # type
+            "1.0.0",  # version
+        )
+        for n in range(n_files)
+    ]
+    t0 = time.time()
+    deps._add_media(values)
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies._add_meta()"
+    t0 = time.time()
+    deps._add_meta("db.new-table.csv", "1.0.0", "archive", "checksum")
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies._drop()"
+    t0 = time.time()
+    deps._drop(["file-9000.wav"])
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies._remove()"
+    t0 = time.time()
+    deps._remove(file)
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = "Dependencies._update_media()"
+    t0 = time.time()
+    deps._update_media(values)
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+    method = f"Dependencies._update_media_version({n_files} files)"
+    t0 = time.time()
+    deps._update_media_version([f"file-{n}.wav" for n in range(n_files)], "version")
+    t = time.time() - t0
+    results.at[method, dtype] = t
+
+# ===== Print results =====
+table = tabulate.tabulate(results, headers="keys", tablefmt="github", floatfmt=".3f")
+print(table)
diff --git a/benchmarks/benchmark-dependency-methods.py b/benchmarks/benchmark-dependency-methods.py
deleted file mode 100644
index e593c978..00000000
--- a/benchmarks/benchmark-dependency-methods.py
+++ /dev/null
@@ -1,209 +0,0 @@
-import hashlib
-import os
-import pickle
-import random
-import string
-import time
-
-import pandas as pd
-
-import audeer
-
-import audb
-
-
-random.seed(1)
-
-cache = audeer.mkdir("./cache")
-
-
-# === Dependency pandas.DataFrame ===
-data_cache = audeer.path(cache, "df.pkl")
-num_rows = 1000000
-if not os.path.exists(data_cache):
-    bit_depths = [0, 16, 24]
-    channels = [0, 1, 2]
-    formats = ["csv", "wav", "txt"]
-    sampling_rates = [0, 16000, 44100]
-    types = [0, 1, 2]
-    versions = ["1.0.0", "1.1.0"]
-    records = [
-        {
-            "file": f"file-{n}.wav",
-            "archive": f"archive-{n}",
-            "bit_depth": random.choices(bit_depths, weights=[0.1, 0.8, 0.1])[0],
-            "channels": random.choices(channels, weights=[0.1, 0.8, 0.1])[0],
-            "checksum": hashlib.md5(
-                pickle.dumps(random.choice(string.ascii_letters))
-            ).hexdigest(),
-            "duration": 10 * random.random(),
-            "format": random.choices(formats, weights=[0.1, 0.8, 0.1])[0],
-            "removed": random.choices([0, 1], weights=[0.1, 0.9])[0],
-            "sampling_rate": random.choices(sampling_rates, weights=[0.1, 0.8, 0.1])[0],
-            "type": random.choices(types, weights=[0.1, 0.8, 0.1])[0],
-            "version": random.choices(versions, weights=[0.2, 0.8])[0],
-        }
-        for n in range(num_rows)
-    ]
-    df = pd.DataFrame.from_records(records)
-    for column, dtype in zip(
-        audb.core.define.DEPEND_FIELD_NAMES.values(),
-        audb.core.define.DEPEND_FIELD_DTYPES.values(),
-    ):
-        df[column] = df[column].astype(dtype)
-    df.set_index("file", inplace=True)
-    df.index.name = None
-    df.index = df.index.astype("string")
-    df.to_pickle(data_cache)
-
-# === Create dependency object ===
-deps = audb.Dependencies()
-deps.load(data_cache)
-
-# ===== Benchmark audb.Dependency =====
-file = "file-10.wav"
-
-t = time.time()
-deps()
-print(f"Dependency.__call__(): {time.time() -t:.3f} s")
-
-# Access the index one time.
-# Further calls will be faster
-file in deps
-
-t = time.time()
-file in deps
-print(f"Dependency.__contains__(): {time.time() -t:.3f} s")
-
-t = time.time()
-deps[file]
-print(f"Dependency.__get_item__(): {time.time() -t:.3f} s")
-
-t = time.time()
-len(deps)
-print(f"Dependency.__len__(): {time.time() -t:.3f} s")
-
-t = time.time()
-str(deps)
-print(f"Dependency.__str__(): {time.time() -t:.3f} s")
-
-t = time.time()
-deps.archives
-print(f"Dependency.archives: {time.time() -t:.3f} s")
-
-t = time.time()
-deps.attachments
-print(f"Dependency.attachments: {time.time() -t:.3f} s")
-
-t = time.time()
-deps.attachment_ids
-print(f"Dependency.attachment_ids: {time.time() -t:.3f} s")
-
-t = time.time()
-files = deps.files
-print(f"Dependency.files: {time.time() -t:.3f} s")
-
-t = time.time()
-deps.media
-print(f"Dependency.media: {time.time() -t:.3f} s")
-
-t = time.time()
-deps.removed_media
-print(f"Dependency.removed_media: {time.time() -t:.3f} s")
-
-t = time.time()
-deps.table_ids
-print(f"Dependency.table_ids: {time.time() -t:.3f} s")
-
-t = time.time()
-deps.tables
-print(f"Dependency.tables: {time.time() -t:.3f} s")
-
-n_files = 1000
-_files = files[:n_files]
-
-t = time.time()
-[deps.archive(file) for file in _files]
-print(f"Dependency.archive({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-[deps.bit_depth(file) for file in _files]
-print(f"Dependency.bit_depth({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-[deps.channels(file) for file in _files]
-print(f"Dependency.channels({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-[deps.checksum(file) for file in _files]
-print(f"Dependency.checksum({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-[deps.duration(file) for file in _files]
-print(f"Dependency.duration({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-[deps.format(file) for file in _files]
-print(f"Dependency.format({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-[deps.removed(file) for file in _files]
-print(f"Dependency.removed({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-[deps.sampling_rate(file) for file in _files]
-print(f"Dependency.sampling_rate({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-[deps.type(file) for file in _files]
-print(f"Dependency.type({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-[deps.version(file) for file in _files]
-print(f"Dependency.version({n_files} files): {time.time() -t:.3f} s")
-
-# -------------------------------------------------------------------------
-t = time.time()
-deps._add_attachment("attachment.txt", "1.0.0", "archive", "checksum")
-print(f"Dependency._add_attachment(): {time.time() -t:.3f} s")
-
-values = [
-    (
-        f"file-new-{n}.wav",  # file
-        f"archive-new-{n}",  # archive
-        16,  # bit_depth
-        1,  # channels
-        f"checksum-{n}",  # checksum
-        0.4,  # duration
-        "wav",  # format
-        0,  # removed
-        16000,  # sampling_rate
-        1,  # type
-        "1.0.0",  # version
-    )
-    for n in range(n_files)
-]
-
-t = time.time()
-deps._add_media(values)
-print(f"Dependency._add_media({n_files} files): {time.time() -t:.3f} s")
-
-t = time.time()
-deps._add_meta("db.new-table.csv", "1.0.0", "archive", "checksum")
-print(f"Dependency._add_meta(): {time.time() -t:.3f} s")
-
-t = time.time()
-deps._drop(["file-9000.wav"])
-print(f"Dependency._drop(): {time.time() -t:.3f} s")
-
-t = time.time()
-deps._remove(file)
-print(f"Dependency._remove(): {time.time() -t:.3f} s")
-
-t = time.time()
-deps._update_media(values)
-print(f"Dependency._update_media(): {time.time() -t:.3f} s")
-
-t = time.time()
-deps._update_media_version([f"file-{n}.wav" for n in range(n_files)], "version")
-print(f"Dependency._update_media_version({n_files} files): {time.time() -t:.3f} s")
diff --git a/benchmarks/requirements.txt b/benchmarks/requirements.txt
new file mode 100644
index 00000000..044a3b3a
--- /dev/null
+++ b/benchmarks/requirements.txt
@@ -0,0 +1,2 @@
+pyarrow
+tabulate