kale-tts-1-tiny / recipe /filter /select_stream.py
NERDDISCO's picture
kale-tts-1
f21d6eb
Raw History Blame Contribute Delete
6.55 kB
# SPDX-License-Identifier: Apache-2.0
# Copyright 2026 Kale
"""Pick the voice designs that pass, as stream_filter.py's rows land (16 Nov 2026; the screen that admitted the newer release voices).
Pass (on the design's own takes, thresholds in config/stream.json "select"): >= 2 of its test-line takes clean (end token, <= 1 missed
word, no clipping, no creak on any sentence end; whispered-by-design voices exempt from creak), test-line pick clean, >= 1 clean take of
its own line, unclear <= 6 per 100 words, UTMOS: test-line pick >= 3.5 (whisper voices >= 3.0), line pick >= 3.0, both picks inside
the pace window (160-200 wpm).
Distinct (ECAPA centroid of the test-line takes): cosine < 0.80 to each voice already in your model (--refs) and < 0.85 to every voice
selected before it (greedy, best first within what has landed).
python3 -I select_stream.py <group> [--refs design/ref-centroids.json] [--once]
Inputs: <root>/design/stream-<group>.jsonl (+ every other design/stream-*.jsonl: those voices count as already selected only if they are
in design/selected.json); --refs = {"<name>": [ECAPA centroid]} for the voices already in your model (e.g. "cent" of their stream rows).
Outputs: <root>/design/selected.json (one row per selected design), the pick wavs copied to <root>/design/selected/<voice>.wav and
<voice>_line.wav, and a decision log <root>/results/select-<group>.log. Loops until stream_filter.py prints DRIVE-DONE into
<root>/logs/drive-<group>.log (or once with --once). root = env KALE_TTS_ROOT or ".".
Trimmed from the original: publishing to our internal listening page (mp3 at -16 LUFS, page manifest, page-specific archive rule)."""
import argparse
import json
import os
import shutil
import tempfile
import time
from pathlib import Path
import numpy as np
ap = argparse.ArgumentParser()
ap.add_argument("group"); ap.add_argument("--refs", default=""); ap.add_argument("--once", action="store_true")
A = ap.parse_args()
H = Path(os.environ.get("KALE_TTS_ROOT", "."))
G, ONCE = A.group, A.once
SEL = {"clean_takes_min": 2, "unclear_max": 6, "utmos_pick_min": 3.5, "utmos_pick_min_whisper": 3.0, "utmos_line_min": 3.0,
"wpm_lo": 160, "wpm_hi": 200, "max_cos_to_model": 0.80, "max_cos_to_selected": 0.85}
for _p in (os.environ.get("KALE_TTS_STREAM_CONFIG"), Path(__file__).resolve().parent / "config" / "stream.json"):
if _p and Path(_p).is_file():
SEL.update(json.loads(Path(_p).read_text()).get("select", {}))
break
REF = {k: np.array(v) for k, v in json.loads(Path(A.refs).read_text()).items()} if A.refs else {}
OUT = H / "design/selected"
OUT.mkdir(parents=True, exist_ok=True)
def write(man):
fd, t = tempfile.mkstemp(dir=H / "design", prefix=".selected.", suffix=".json")
with os.fdopen(fd, "w") as fh: json.dump(man, fh, indent=1)
os.replace(t, H / "design/selected.json")
def passed(o):
s, ex = o["scores"], o.get("creak_exempt")
lp = o.get("line_pick") or {}
return (o["clean_takes"] >= SEL["clean_takes_min"] and o["pick"]["clean"] and o["line_clean"] >= 1 and lp.get("clean") and s["unclear"] <= SEL["unclear_max"]
and o["pick"]["utmos"] >= (SEL["utmos_pick_min_whisper"] if ex else SEL["utmos_pick_min"]) and lp.get("utmos", 0) >= SEL["utmos_line_min"]
and SEL["wpm_lo"] <= o["pick"]["wpm_full"] <= SEL["wpm_hi"] and SEL["wpm_lo"] <= lp.get("wpm_full", 0) <= SEL["wpm_hi"])
(H / "results").mkdir(parents=True, exist_ok=True)
log = open(H / f"results/select-{G}.log", "a")
seen = set()
while True:
sp = H / f"design/stream-{G}.jsonl"
rows = [json.loads(l) for l in sp.read_text().splitlines()] if sp.exists() else []
man = json.loads((H / "design/selected.json").read_text()) if (H / "design/selected.json").exists() else []
have = {m["id"] for m in man}
E = {}
for f in H.glob("design/stream-*.jsonl"): # every streamed design (other groups too)
for l in f.read_text().splitlines(): E[json.loads(l)["voice"]] = np.array(json.loads(l)["cent"])
on_list = [E[i] for i in have if i in E]
new = [r for r in rows if r["voice"] not in seen and r["voice"] not in have]
new.sort(key=lambda o: -(o["scores"]["utmos"] + o["clean_takes"] / max(1, o["takes"]) + 0.3 * (o.get("line_pick") or {}).get("utmos", 0)))
for o in new:
seen.add(o["voice"])
c = E[o["voice"]]
nc = max(((k, float(c @ REF[k])) for k in REF), key=lambda x: x[1], default=("-", 0.0))
na = max([float(c @ x) for x in on_list] or [0])
why = None if passed(o) else "quality"
if not why and nc[1] >= SEL["max_cos_to_model"]: why = f"close to model voice {nc[0]} {nc[1]:.3f}"
if not why and na >= SEL["max_cos_to_selected"]: why = f"close to a selected voice {na:.3f}"
log.write(json.dumps(dict(t=time.strftime("%H:%M"), voice=o["voice"], label=o.get("label"), why=why or "selected", clean=o["clean_takes"], line_clean=o["line_clean"],
scores=o["scores"], pick_utmos=o["pick"]["utmos"], line_utmos=(o.get("line_pick") or {}).get("utmos"))) + "\n"); log.flush()
if why:
continue
for src, f in ((o["raw"], f"{o['voice']}.wav"), (o["raw_line"], f"{o['voice']}_line.wav")):
shutil.copyfile(src, OUT / f)
man.append(dict(id=o["voice"], n=len(man) + 1, label=o.get("label"), gender=o["gender"], age=o.get("age"), accent=o.get("accent"),
wpm=round(o["pick"]["wpm_full"]), target_wpm=o["target_wpm"], file=f"selected/{o['voice']}.wav", line_text=o["line"],
line_file=f"selected/{o['voice']}_line.wav", line_wpm=round(o["line_pick"]["wpm_full"]), description=o["description"],
scores=o["scores"], clean_takes=f"{o['clean_takes']}/{o['takes']}", pick_utmos=o["pick"]["utmos"], line_utmos=o["line_pick"]["utmos"],
nearest_model_voice=dict(voice=nc[0], cos=round(nc[1], 3)), nearest_selected_cos=round(na, 3),
source=f"original Maya-1 audio (designed voice, vLLM bf16), takes {o['pick']['clip']} + {o['line_pick']['clip']}"))
on_list.append(c); have.add(o["voice"])
write(man)
print(time.strftime("%H:%M"), "selected", o["voice"], o.get("label"), "| list now", len(man), flush=True)
done = "DRIVE-DONE" in (H / f"logs/drive-{G}.log").read_text() if (H / f"logs/drive-{G}.log").exists() else False
if ONCE or (done and not [r for r in rows if r["voice"] not in seen]):
break
time.sleep(20)
print("SELECT-DONE", G, flush=True)