add speaker prompts

libritts-r-aligned.py

@@ -15,6 +15,7 @@ from multiprocessing import cpu_count
 from phones.convert import Converter
 import torchaudio
 import torchaudio.transforms as AT
+from functools import lru_cache
 
 logger = datasets.logging.get_logger(__name__)
 
@@ -64,6 +65,12 @@ _URLS = {
     "train-other-500": _URL + "train_other_500.tar.gz",
 }
 
+@lru_cache(maxsize=1000)
+def get_speaker_prompts(speaker, hash_ds):
+    ds = hash_ds.df
+    speaker_prompts = ds[ds["speaker"] == speaker]
+    speaker_prompts = tuple(speaker_prompts["audio"])
+    return speaker_prompts
 
 class LibriTTSAlignConfig(datasets.BuilderConfig):
     """BuilderConfig for LibriTTSAlign."""
@@ -106,7 +113,8 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
             "phones": datasets.Sequence(datasets.Value("string")),
             "phone_durations": datasets.Sequence(datasets.Value("int32")),
             # audio feature
-            "audio": datasets.Value("string")
+            "audio": datasets.Value("string"),
+            "audio_speaker_prompt": datasets.Sequence(datasets.Value("string")),
         }
 
         return datasets.DatasetInfo(
@@ -159,19 +167,25 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
         speakers = data_all["speaker"].unique()
         # seed for reproducibility
         np.random.seed(42)
-        data_dev_all = None
-        for speaker in tqdm(speakers, desc="creating dev split"):
-            data_speaker = data_all[data_all["speaker"] == speaker]
-            if len(data_speaker) < 10:
-                print(f"Speaker {speaker} has only {len(data_speaker)} samples, skipping")
-            else:
-                data_speaker = data_speaker.sample(1)
-                data_all = data_all[data_all["audio"] != data_speaker["audio"].values[0]]
-                if data_dev_all is None:
-                    data_dev_all = data_speaker
-                else:
-                    data_dev_all = pd.concat([data_dev_all, data_speaker])
+        self.data_all = data_all
+        del data_all
+        data_dev_all = [
+            x for x in 
+            process_map(
+                self._create_dev_split,
+                speakers,
+                chunksize=1000,
+                max_workers=_MAX_WORKERS,
+                desc="creating dev split",
+                tqdm_class=tqdm,
+            )
+            if x is not None
+        ]
+        data_dev_all = pd.concat(data_dev_all)
+        data_all = self.data_all
         data_all = data_all[data_all["speaker"].isin(data_dev_all["speaker"].unique())]
+        data_all = data_all[~data_all["basename"].isin(data_dev_all["basename"].unique())]
+        del self.data_all
         self.speaker2idxs = {}
         self.speaker2idxs["all"] = {speaker: idx for idx, speaker in enumerate(sorted(list(data_dev_all["speaker"].unique())))}
         self.speaker2idxs["train"] = {speaker: idx for idx, speaker in enumerate(sorted(list(data_train["speaker"].unique())))}
@@ -194,6 +208,15 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
         self.alignments_ds = None
         self.data = None
         return splits
+    
+    def _create_dev_split(self, speaker):
+        data_speaker = self.data_all[self.data_all["speaker"] == speaker]
+        if len(data_speaker) < 10:
+            print(f"Speaker {speaker} has only {len(data_speaker)} samples, skipping")
+            return None
+        else:
+            data_speaker = data_speaker.sample(2)
+            return data_speaker
 
     def _create_alignments_ds(self, name, url):
         self.empty_textgrids = 0
@@ -251,7 +274,7 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
         if self.empty_textgrids > 0:
             logger.warning(f"Found {self.empty_textgrids} empty textgrids")
         del self.ds, self.phone_cache, self.phone_converter
-        return pd.DataFrame(
+        df = pd.DataFrame(
             entries,
             columns=[
                 "phones",
@@ -264,6 +287,7 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
                 "basename",
             ],
         )
+        return df
 
     def _create_entry(self, dsi_idx):
         dsi, idx = dsi_idx
@@ -298,7 +322,7 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
         if start >= end:
             self.empty_textgrids += 1
             return None
-
+        
         return (
             phones,
             durations,
@@ -312,10 +336,13 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
 
     def _generate_examples(self, ds):
         j = 0
+        hash_col = "audio"
+        hash_ds = HashableDataFrame(ds, hash_col)
         for i, row in ds.iterrows():
             # 10kB is the minimum size of a wav file for our purposes
             if Path(row["audio"]).stat().st_size >= 10_000:
                 if len(row["phones"]) < 384:
+                    speaker_prompts = get_speaker_prompts(row["speaker"], hash_ds)
                     result = {
                         "id": row["basename"],
                         "speaker": row["speaker"],
@@ -325,6 +352,21 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
                         "phones": row["phones"],
                         "phone_durations": row["duration"],
                         "audio": str(row["audio"]),
+                        "audio_speaker_prompt": speaker_prompts,
                     }
                     yield j, result
-                    j += 1
+                    j += 1
+
+class HashableDataFrame():
+    def __init__(self, df, hash_col):
+        self.df = df
+        self.hash_col = hash_col
+        self.hash = hashlib.md5(self.df[self.hash_col].values).hexdigest()
+        # to integer
+        self.hash = int(self.hash, 16)
+    
+    def __hash__(self):
+        return self.hash
+
+    def __eq__(self, other):
+        return self.hash == other.hash