File size: 3,486 Bytes
1a3acb4 e72bdde 1a3acb4 e72bdde 1a3acb4 e72bdde 1a3acb4 e72bdde 1a3acb4 e72bdde | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 | """emo_data_prep.py v2 — R3 emotion-tagged training data.
Rows: "Speaker 0: <emotion> {transcription}" with audio = the acted clip (24 kHz) and
voice_prompts = [a DIFFERENT clip of the SAME actor+emotion] (list form, matching DailyTalk's
schema). Mixed with a DailyTalk sample (untagged) to prevent forgetting.
Outputs: /data/mixed_train.jsonl, /data/mixed_val.jsonl. Prints EMO PREP DONE.
"""
import json, os, random, sys
import numpy as np
import soundfile as sf
import librosa
SR = 24000
TAG = {"ANGRY": "<angry>", "SAD": "<sad>", "HAPPY": "<happy>",
"NEUTRAL": "<neutral>", "DISGUST": "<disgust>", "FEARFUL": "<fearful>"}
CAP_PER_EMOTION = 250
DT_SAMPLE = 300
VAL_EMOTION = 45
VAL_DT = 15
random.seed(1234)
def main():
from datasets import load_dataset
ds = load_dataset("zenitsu09/CREMAD-emotion-transcribed", split="train")
print(f"CREMA-D loaded: {len(ds)} rows", flush=True)
os.makedirs("/emo/audio", exist_ok=True)
rows = []
count = {}
n_emo = 0
for row in ds:
emo_raw = str(row["emotion"]).upper()
emo = "ANGRY" if emo_raw.startswith("ANG") else emo_raw
if emo not in TAG:
continue
if count.get(emo, 0) >= CAP_PER_EMOTION:
continue
count[emo] = count.get(emo, 0) + 1
arr = np.asarray(row["audio"]["array"], dtype=np.float32)
sr = row["audio"]["sampling_rate"]
if sr != SR:
arr = librosa.resample(arr, orig_sr=sr, target_sr=SR)
text = str(row["transcription"]).strip()
if text.startswith("<"):
text = text.split(">", 1)[-1].strip()
if not text or len(text) < 4:
continue
n_emo += 1
fn = f"/emo/audio/e{n_emo:05d}_{emo.lower()}.wav"
sf.write(fn, arr.astype(np.float32), SR)
actor = str(row["original_filename"]).split("_")[0]
rows.append({"text": f"Speaker 0: {TAG[emo]} {text}", "audio": fn,
"actor": actor, "emo": emo})
print(f"emotion rows: {n_emo} ({count})", flush=True)
groups = {}
for r in rows:
groups.setdefault((r["actor"], r["emo"]), []).append(r)
paired = 0
for r in rows:
pool = [x for x in groups.get((r["actor"], r["emo"]), []) if x is not r]
if pool:
r["voice_prompts"] = [random.choice(pool)["audio"]]
paired += 1
else:
r["voice_prompts"] = [r["audio"]]
print(f"paired with different same-actor-same-emotion clip: {paired}/{n_emo}", flush=True)
dt = [json.loads(l) for l in open("/data/train.jsonl")]
random.shuffle(dt)
dt_sample = dt[:DT_SAMPLE]
print(f"dailytalk sample: {len(dt_sample)}", flush=True)
assert all(isinstance(r.get("voice_prompts"), list) for r in dt_sample[:5]), "dt schema check"
emo_rows = [{"text": r["text"], "audio": r["audio"], "voice_prompts": r["voice_prompts"]}
for r in rows]
random.shuffle(emo_rows)
val = emo_rows[:VAL_EMOTION] + dt_sample[:VAL_DT]
train = emo_rows[VAL_EMOTION:] + dt_sample[VAL_DT:]
random.shuffle(train)
random.shuffle(val)
with open("/data/mixed_train.jsonl", "w") as f:
for r in train:
f.write(json.dumps(r) + "\n")
with open("/data/mixed_val.jsonl", "w") as f:
for r in val:
f.write(json.dumps(r) + "\n")
print(f"MIXED TRAIN {len(train)} / VAL {len(val)}", flush=True)
print("EMO PREP DONE", flush=True)
if __name__ == "__main__":
main() |