Commit
·
6b456c1
1
Parent(s):
b1bc032
save buttons bro
Browse files
utils.py
CHANGED
@@ -237,3 +237,93 @@ def wav_bytes_base64(x: np.ndarray, sr: int) -> tuple[str, int, int]:
|
|
237 |
buf.seek(0)
|
238 |
b64 = base64.b64encode(buf.read()).decode("utf-8")
|
239 |
return b64, int(x.shape[0]), int(x.shape[1])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
237 |
buf.seek(0)
|
238 |
b64 = base64.b64encode(buf.read()).decode("utf-8")
|
239 |
return b64, int(x.shape[0]), int(x.shape[1])
|
240 |
+
|
241 |
+
def _ratio(out_sr: int, in_sr: int) -> tuple[int, int]:
|
242 |
+
g = gcd(int(out_sr), int(in_sr))
|
243 |
+
return int(out_sr) // g, int(in_sr) // g
|
244 |
+
|
245 |
+
class StreamingResampler:
|
246 |
+
"""
|
247 |
+
Stateful streaming resampler.
|
248 |
+
Prefers soxr (best), then libsamplerate; final fallback is block resample_poly.
|
249 |
+
Always pass float32 arrays shaped (S, C).
|
250 |
+
"""
|
251 |
+
def __init__(self, in_sr: int, out_sr: int, channels: int = 2, quality: str = "VHQ"):
|
252 |
+
self.in_sr = int(in_sr)
|
253 |
+
self.out_sr = int(out_sr)
|
254 |
+
self.channels = int(channels)
|
255 |
+
self.quality = quality
|
256 |
+
self._backend = None
|
257 |
+
|
258 |
+
# Try soxr first
|
259 |
+
try:
|
260 |
+
import soxr # pip install soxr
|
261 |
+
self._backend = "soxr"
|
262 |
+
# dtype float32 keeps things consistent with the rest of your code
|
263 |
+
self._rs = soxr.Resampler(
|
264 |
+
self.in_sr,
|
265 |
+
self.out_sr,
|
266 |
+
channels=self.channels,
|
267 |
+
dtype="float32",
|
268 |
+
quality=self.quality, # "Q", "HQ", "VHQ"
|
269 |
+
)
|
270 |
+
except Exception:
|
271 |
+
# Try libsamplerate
|
272 |
+
try:
|
273 |
+
import samplerate # pip install samplerate
|
274 |
+
self._backend = "samplerate"
|
275 |
+
# sinc_best == highest quality; you can choose 'sinc_medium' for speed
|
276 |
+
self._rs = samplerate.Resampler(converter_type="sinc_best", channels=self.channels)
|
277 |
+
except Exception:
|
278 |
+
# Last resort: block resample (not truly streaming)
|
279 |
+
from scipy.signal import resample_poly
|
280 |
+
self._backend = "scipy"
|
281 |
+
self._resample_poly = resample_poly
|
282 |
+
self._L, self._M = _ratio(self.out_sr, self.in_sr)
|
283 |
+
# Keep a tiny tail to help transitions (still not perfect vs true streaming)
|
284 |
+
self._hist = np.zeros((0, self.channels), dtype=np.float32)
|
285 |
+
|
286 |
+
def process(self, x: np.ndarray, final: bool = False) -> np.ndarray:
|
287 |
+
"""Feed a chunk (S, C) and get resampled chunk (S', C). Keep calling in order."""
|
288 |
+
if x.size == 0 and not final:
|
289 |
+
# nothing to do
|
290 |
+
return np.zeros((0, self.channels), dtype=np.float32)
|
291 |
+
|
292 |
+
if self._backend == "soxr":
|
293 |
+
return self._rs.process(x, final=final)
|
294 |
+
|
295 |
+
elif self._backend == "samplerate":
|
296 |
+
import samplerate
|
297 |
+
ratio = float(self.out_sr) / float(self.in_sr)
|
298 |
+
# end_of_input=True flushes tail on the last call
|
299 |
+
y = self._rs.process(x, ratio, end_of_input=final)
|
300 |
+
# libsamplerate returns (S', C)
|
301 |
+
return y.astype(np.float32, copy=False)
|
302 |
+
|
303 |
+
# --- scipy fallback (block, not truly streaming) ---
|
304 |
+
# We concatenate a short history to reduce block edge artifacts
|
305 |
+
x_ext = x if self._hist.size == 0 else np.vstack([self._hist, x])
|
306 |
+
y = self._resample_poly(x_ext, up=self._L, down=self._M, axis=0).astype(np.float32, copy=False)
|
307 |
+
|
308 |
+
# Heuristic: drop the portion corresponding roughly to the history to avoid duplicate content
|
309 |
+
# (Not perfect, but helps a lot when chunks are reasonably sized.)
|
310 |
+
drop = int(round(self._hist.shape[0] * self.out_sr / self.in_sr))
|
311 |
+
y = y[drop:] if drop < y.shape[0] else np.zeros((0, self.channels), dtype=np.float32)
|
312 |
+
|
313 |
+
# Keep a small input tail for the next call (say ~ 4 ms at in_sr)
|
314 |
+
tail_samples = max(int(0.004 * self.in_sr), 1)
|
315 |
+
self._hist = x[-tail_samples:] if x.shape[0] >= tail_samples else x.copy()
|
316 |
+
if final:
|
317 |
+
self._hist = np.zeros((0, self.channels), dtype=np.float32)
|
318 |
+
return y
|
319 |
+
|
320 |
+
def flush(self) -> np.ndarray:
|
321 |
+
"""Drain converter tail (call at stop)."""
|
322 |
+
if self._backend == "soxr":
|
323 |
+
return self._rs.process(np.zeros((0, self.channels), dtype=np.float32), final=True)
|
324 |
+
elif self._backend == "samplerate":
|
325 |
+
ratio = float(self.out_sr) / float(self.in_sr)
|
326 |
+
return self._rs.process(np.zeros((0, self.channels), dtype=np.float32), ratio, end_of_input=True)
|
327 |
+
else:
|
328 |
+
# nothing meaningful to flush in scipy fallback
|
329 |
+
return np.zeros((0, self.channels), dtype=np.float32)
|