Initial commit: RecorderStudio：PyQt5 高保真录音软件，无损 WAV / WASAPI 独占、BS.1770 响度与 ffmpeg 交叉验证
This commit is contained in:
WpyQwq
@@ -0,0 +1,579 @@
|
||||
"""无损 WAV 读写核心。
|
||||
|
||||
设计要点（面向"高质量、高清晰度"）：
|
||||
|
||||
* 支持 16 / 24 / 32 位整数 PCM 与 32 位浮点（IEEE float）四种载荷；
|
||||
* 24 位采用真正的 3 字节打包，不做 16 位截断；
|
||||
* 多声道、24 位、浮点载荷统一使用 ``WAVE_FORMAT_EXTENSIBLE``（含通道掩码），
|
||||
避免播放器把 24 位误判为 16 位、把浮点误判为整数；
|
||||
* 高位深写入时可选 **TPDF（三角概率密度）抖动**，把量化失真从
|
||||
"与信号相关的非线性失真" 变成 "与信号无关的宽带白噪"，这是后期降位深的
|
||||
标准做法，听感明显更干净；
|
||||
* 录音过程中定期回写文件头（``checkpoint``），即使程序崩溃/断电，
|
||||
已落盘的数据也能被播放器正常识别；
|
||||
* 超过 4 GiB 时可选 RF64（EBU Tech 3306）容器，避免 RIFF 32 位长度溢出。
|
||||
|
||||
本模块不依赖 numpy 以外的任何第三方库。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
import struct
|
||||
from dataclasses import dataclass
|
||||
from typing import BinaryIO, Iterator
|
||||
|
||||
import numpy as np
|
||||
|
||||
WAVE_FORMAT_PCM = 0x0001
|
||||
WAVE_FORMAT_IEEE_FLOAT = 0x0003
|
||||
WAVE_FORMAT_EXTENSIBLE = 0xFFFE
|
||||
|
||||
# KSDATAFORMAT_SUBTYPE_*
|
||||
_SUBTYPE_PCM = bytes.fromhex("0100000000001000800000aa00389b71")
|
||||
_SUBTYPE_FLOAT = bytes.fromhex("0300000000001000800000aa00389b71")
|
||||
|
||||
# 声道掩码（dwChannelMask），用于 EXTENSIBLE 头部
|
||||
_CHANNEL_MASKS = {
|
||||
1: 0x4, # SPEAKER_FRONT_CENTER -> 单声道习惯上用 CENTER，但很多软件写 FRONT_LEFT
|
||||
2: 0x3, # FRONT_LEFT | FRONT_RIGHT
|
||||
4: 0x33, # FL FR BL BR
|
||||
6: 0x3F,
|
||||
8: 0x63F,
|
||||
}
|
||||
|
||||
# 1 GiB 保留量：RIFF 数据块上限 4 GiB-1，留出余量给头部与安全边界
|
||||
RIFF_DATA_LIMIT = 0xFFF00000 - 64
|
||||
|
||||
__all__ = [
|
||||
"WavFormat",
|
||||
"WavWriter",
|
||||
"WavReader",
|
||||
"wav_info",
|
||||
"read_wav",
|
||||
"write_wav",
|
||||
"RIFF_DATA_LIMIT",
|
||||
"SUPPORTED_BIT_DEPTHS",
|
||||
]
|
||||
|
||||
# (名称, 位深, 采样格式) —— 直接对应界面上的"位深"下拉框
|
||||
SUPPORTED_BIT_DEPTHS = ("16", "24", "32", "float32")
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class WavFormat:
|
||||
"""描述一个 WAV 载荷格式。"""
|
||||
|
||||
samplerate: int
|
||||
channels: int
|
||||
encoding: str # 'pcm' | 'float'
|
||||
bits: int # 每个采样的有效位宽（float 为 32）
|
||||
|
||||
@property
|
||||
def bytes_per_sample(self) -> int:
|
||||
return self.bits // 8
|
||||
|
||||
@property
|
||||
def frame_bytes(self) -> int:
|
||||
return self.bytes_per_sample * self.channels
|
||||
|
||||
@property
|
||||
def bytes_per_second(self) -> float:
|
||||
return float(self.frame_bytes) * self.samplerate
|
||||
|
||||
@property
|
||||
def block_align(self) -> int:
|
||||
return self.frame_bytes
|
||||
|
||||
@property
|
||||
def label(self) -> str:
|
||||
if self.encoding == "float":
|
||||
return "32-bit float"
|
||||
return f"{self.bits}-bit PCM"
|
||||
|
||||
def describe(self) -> str:
|
||||
return (
|
||||
f"{self.samplerate} Hz / {self.label} / "
|
||||
f"{'单声道' if self.channels == 1 else f'{self.channels} 声道'}"
|
||||
)
|
||||
|
||||
|
||||
def _fmt_for_bitdepth(bit_depth: int | str) -> tuple[str, int]:
|
||||
"""把界面上的位深选项翻译成 (encoding, bits)。"""
|
||||
if isinstance(bit_depth, str):
|
||||
s = bit_depth.strip().lower()
|
||||
if s in ("float", "float32", "f32", "32f"):
|
||||
return "float", 32
|
||||
s = s.rstrip("bit").rstrip("-").strip()
|
||||
bit_depth = int(s)
|
||||
bit_depth = int(bit_depth)
|
||||
if bit_depth == 16:
|
||||
return "pcm", 16
|
||||
if bit_depth == 24:
|
||||
return "pcm", 24
|
||||
if bit_depth == 32:
|
||||
return "pcm", 32
|
||||
if bit_depth == 64:
|
||||
raise ValueError("不支持 64 位整数 PCM；如需更高精度请使用 32-bit float")
|
||||
raise ValueError(f"不支持的位深: {bit_depth}")
|
||||
|
||||
|
||||
def _safe_scale(bits: int) -> float:
|
||||
"""整数满量程对应的浮点值。
|
||||
|
||||
采用 ``2**(bits-1) - 1``（对称满量程），这样 +1.0 与 -1.0 都能被精确表示，
|
||||
不会出现 "正半周先削顶、负半周仍有余量" 的非对称削波。
|
||||
"""
|
||||
return float((1 << (bits - 1)) - 1)
|
||||
|
||||
|
||||
class _TpdfDither:
|
||||
"""TPDF 抖动噪声源。
|
||||
|
||||
三角分布 = 两个独立均匀分布之和，峰峰值恰好 1 LSB，
|
||||
这是业界公认的"无调制噪声"抖动，优于矩形（RPDF）抖动。
|
||||
"""
|
||||
|
||||
def __init__(self, seed: int | None = None):
|
||||
self._rng = np.random.Generator(np.random.PCG64(seed))
|
||||
|
||||
def generate(self, size: int | tuple[int, ...], lsb: float = 1.0) -> np.ndarray:
|
||||
u1 = self._rng.random(size)
|
||||
u2 = self._rng.random(size)
|
||||
return (u1 - u2) * lsb
|
||||
|
||||
def reseed(self, seed: int) -> None:
|
||||
self._rng = np.random.Generator(np.random.PCG64(seed))
|
||||
|
||||
|
||||
def float_to_pcm(data: np.ndarray, encoding: str, bits: int,
|
||||
dither: _TpdfDither | None = None) -> bytes:
|
||||
"""把 float32/float64 的 [-1, 1] 数据编码为 WAV 载荷字节。"""
|
||||
x = np.asarray(data)
|
||||
if encoding == "float":
|
||||
return np.ascontiguousarray(x, dtype="<f4").tobytes()
|
||||
|
||||
# 抖动与量化都必须在"整数域"内进行：抖动幅度 = 1 LSB（峰峰值）
|
||||
scale = _safe_scale(bits)
|
||||
y = x.astype(np.float64, copy=True) * scale
|
||||
if dither is not None:
|
||||
y += dither.generate(y.shape, lsb=1.0)
|
||||
|
||||
ints = np.rint(y)
|
||||
lo, hi = -(1 << (bits - 1)), (1 << (bits - 1)) - 1
|
||||
np.clip(ints, lo, hi, out=ints)
|
||||
ints = ints.astype(np.int64)
|
||||
|
||||
if bits == 16:
|
||||
return ints.astype("<i2").tobytes()
|
||||
if bits == 32:
|
||||
return ints.astype("<i4").tobytes()
|
||||
if bits == 24:
|
||||
buf = np.ascontiguousarray(ints, dtype="<i4")
|
||||
# 小端序下低 3 字节即为 24 位补码，直接切片比逐样本运算快两个数量级
|
||||
return buf.view(np.uint8).reshape(-1, 4)[:, :3].tobytes()
|
||||
raise ValueError(f"不支持的位深: {bits}")
|
||||
|
||||
|
||||
def pcm_to_float(raw: bytes, encoding: str, bits: int) -> np.ndarray:
|
||||
"""WAV 载荷字节 -> float64 数组（[-1, 1] 标度）。"""
|
||||
if encoding == "float":
|
||||
return np.frombuffer(raw, dtype="<f4").astype(np.float64)
|
||||
if bits == 16:
|
||||
return np.frombuffer(raw, dtype="<i2").astype(np.float64) / _safe_scale(16)
|
||||
if bits == 32:
|
||||
return np.frombuffer(raw, dtype="<i4").astype(np.float64) / _safe_scale(32)
|
||||
if bits == 24:
|
||||
b = np.frombuffer(raw, dtype=np.uint8)
|
||||
usable = (b.size // 3) * 3
|
||||
b = b[:usable].reshape(-1, 3).astype(np.int32)
|
||||
v = b[:, 0] | (b[:, 1] << 8) | (b[:, 2] << 16)
|
||||
# 手动符号扩展
|
||||
v = np.where(v & 0x800000, v - 0x1000000, v)
|
||||
return v.astype(np.float64) / _safe_scale(24)
|
||||
raise ValueError(f"不支持的位深: {bits}")
|
||||
|
||||
|
||||
class WavWriter:
|
||||
"""流式（边录边写）WAV 写入器。
|
||||
|
||||
典型用法::
|
||||
|
||||
w = WavWriter("a.wav", WavFormat(48000, 2, "pcm", 24), dither=True)
|
||||
w.write(block) # block: float32 (n, channels)
|
||||
w.checkpoint() # 每秒调用一次，保证崩溃安全
|
||||
info = w.close() # 回写文件头，返回统计信息
|
||||
"""
|
||||
|
||||
def __init__(self, path: str, fmt: WavFormat, *, dither: bool = True,
|
||||
rf64: bool = False, seed: int | None = None,
|
||||
checkpoint_interval: float = 1.0):
|
||||
self.path = os.fspath(path)
|
||||
self.fmt = fmt
|
||||
self.rf64 = bool(rf64)
|
||||
self._dither = _TpdfDither(seed) if (dither and fmt.encoding == "pcm") else None
|
||||
self._checkpoint_interval = float(checkpoint_interval)
|
||||
self._last_checkpoint = 0.0
|
||||
|
||||
self.frames_written = 0
|
||||
self.bytes_written = 0
|
||||
self.clipped_samples = 0
|
||||
self.peak = 0.0
|
||||
self.closed = False
|
||||
|
||||
self._extensible = (
|
||||
fmt.encoding == "float" or fmt.bits > 16 or fmt.channels > 2
|
||||
)
|
||||
self._fh: BinaryIO = open(self.path, "wb", buffering=1024 * 1024)
|
||||
hdr = self._build_header()
|
||||
self._header = hdr
|
||||
self._data_size_offset = hdr.data_offset - 4
|
||||
self._fh.write(hdr.bytes)
|
||||
self._fh.flush()
|
||||
|
||||
# ---------------------------------------------------------------- header
|
||||
def _build_header(self) -> "_Header":
|
||||
f = self.fmt
|
||||
if f.encoding == "float":
|
||||
fmt_tag = WAVE_FORMAT_IEEE_FLOAT if not self._extensible else WAVE_FORMAT_EXTENSIBLE
|
||||
subtype = _SUBTYPE_FLOAT
|
||||
else:
|
||||
fmt_tag = WAVE_FORMAT_PCM if not self._extensible else WAVE_FORMAT_EXTENSIBLE
|
||||
subtype = _SUBTYPE_PCM
|
||||
|
||||
if self._extensible:
|
||||
mask = _CHANNEL_MASKS.get(f.channels, 0)
|
||||
fmt_body = struct.pack(
|
||||
"<HHIIHHHHI16s",
|
||||
WAVE_FORMAT_EXTENSIBLE, f.channels, f.samplerate,
|
||||
int(f.bytes_per_second), f.block_align, f.bits,
|
||||
22, f.bits, mask, subtype,
|
||||
)
|
||||
else:
|
||||
fmt_body = struct.pack(
|
||||
"<HHIIHH",
|
||||
fmt_tag, f.channels, f.samplerate,
|
||||
int(f.bytes_per_second), f.block_align, f.bits,
|
||||
)
|
||||
|
||||
chunks: list[bytes] = []
|
||||
# fmt
|
||||
chunks.append(b"fmt " + struct.pack("<I", len(fmt_body)) + fmt_body)
|
||||
# fact（非 PCM 载荷必须声明采样帧数）
|
||||
if f.encoding == "float":
|
||||
chunks.append(b"fact" + struct.pack("<II", 4, 0))
|
||||
self._fact_offset = 12 + len(chunks[0]) + 8
|
||||
else:
|
||||
self._fact_offset = None
|
||||
|
||||
riff_overhead = 12 + sum(len(c) for c in chunks) + 8
|
||||
if self.rf64:
|
||||
# RF64: ds64 紧跟 WAVE 之后，RIFF size 写 0xFFFFFFFF
|
||||
ds64_body = struct.pack("<QQQI", 0, 0, 0, 0)
|
||||
ds64 = b"ds64" + struct.pack("<I", len(ds64_body)) + ds64_body
|
||||
body = ds64 + b"".join(chunks)
|
||||
header = (b"RF64" + struct.pack("<I", 0xFFFFFFFF) + b"WAVE"
|
||||
+ body + b"data" + struct.pack("<I", 0))
|
||||
return _Header(header, len(header), riff_overhead, ds64_offset=12)
|
||||
body = b"".join(chunks)
|
||||
size = len(body) + 8 + 8 # 占位，收尾时回写
|
||||
header = (b"RIFF" + struct.pack("<I", size) + b"WAVE" + body
|
||||
+ b"data" + struct.pack("<I", 0))
|
||||
return _Header(header, len(header), riff_overhead, ds64_offset=None)
|
||||
|
||||
# ----------------------------------------------------------------- write
|
||||
def write(self, data: np.ndarray) -> int:
|
||||
"""写入一块浮点音频（形状 (n,) 或 (n, channels)），返回写入的采样帧数。"""
|
||||
if self.closed:
|
||||
raise ValueError("写入器已关闭")
|
||||
x = np.asarray(data, dtype=np.float32)
|
||||
if x.ndim == 1:
|
||||
x = x.reshape(-1, 1)
|
||||
if x.shape[1] != self.fmt.channels:
|
||||
raise ValueError(
|
||||
f"通道数不匹配：收到 {x.shape[1]}，期望 {self.fmt.channels}")
|
||||
|
||||
n = int(x.shape[0])
|
||||
if n == 0:
|
||||
return 0
|
||||
|
||||
# 统计（在抖动/削波之前，反映真实输入）
|
||||
blk_peak = float(np.max(np.abs(x))) if n else 0.0
|
||||
if blk_peak > self.peak:
|
||||
self.peak = blk_peak
|
||||
if blk_peak > 1.0:
|
||||
self.clipped_samples += int(np.count_nonzero(np.abs(x) > 1.0))
|
||||
|
||||
payload = float_to_pcm(x, self.fmt.encoding, self.fmt.bits, self._dither)
|
||||
|
||||
# 先判断是否越界，再落盘：这样调用方可以无损地切到下一个文件
|
||||
if not self.rf64 and self.bytes_written + len(payload) > RIFF_DATA_LIMIT:
|
||||
raise _RiffOverflow(
|
||||
f"RIFF 容器已达 4 GiB 上限（已写 {self.bytes_written} 字节），"
|
||||
"请启用自动分段或 RF64 模式")
|
||||
|
||||
self._fh.write(payload)
|
||||
self.frames_written += n
|
||||
self.bytes_written += len(payload)
|
||||
return n
|
||||
|
||||
def checkpoint(self, now: float | None = None) -> None:
|
||||
"""定期回写长度字段，保证异常退出后文件仍可播放。"""
|
||||
if self.closed:
|
||||
return
|
||||
now = _monotonic() if now is None else now
|
||||
if now - self._last_checkpoint < self._checkpoint_interval:
|
||||
return
|
||||
self._last_checkpoint = now
|
||||
self._patch_sizes(final=False)
|
||||
|
||||
def _patch_sizes(self, *, final: bool) -> None:
|
||||
f = self.fmt
|
||||
data_bytes = self.frames_written * f.frame_bytes
|
||||
pos = self._fh.tell()
|
||||
if self.rf64:
|
||||
body_len = len(self._header.bytes) - 8 + data_bytes
|
||||
if final:
|
||||
self._fh.seek(4)
|
||||
self._fh.write(struct.pack("<I", 0xFFFFFFFF))
|
||||
self._fh.seek(self._header.ds64_offset + 8) # ds64 body 起点
|
||||
self._fh.write(struct.pack(
|
||||
"<QQQI", body_len, data_bytes, self.frames_written, 0))
|
||||
self._fh.seek(pos)
|
||||
return
|
||||
# RIFF
|
||||
riff_size = (len(self._header.bytes) - 8) + data_bytes
|
||||
self._fh.seek(4)
|
||||
self._fh.write(struct.pack("<I", riff_size & 0xFFFFFFFF))
|
||||
self._fh.seek(self._data_size_offset)
|
||||
self._fh.write(struct.pack("<I", data_bytes & 0xFFFFFFFF))
|
||||
if self._fact_offset is not None:
|
||||
self._fh.seek(self._fact_offset)
|
||||
self._fh.write(struct.pack("<I", self.frames_written & 0xFFFFFFFF))
|
||||
self._fh.seek(pos)
|
||||
|
||||
# ----------------------------------------------------------------- close
|
||||
def close(self) -> dict:
|
||||
if self.closed:
|
||||
return self.stats()
|
||||
self._patch_sizes(final=True)
|
||||
self._fh.flush()
|
||||
try:
|
||||
os.fsync(self._fh.fileno())
|
||||
except OSError:
|
||||
pass
|
||||
self._fh.close()
|
||||
self.closed = True
|
||||
return self.stats()
|
||||
|
||||
def stats(self) -> dict:
|
||||
f = self.fmt
|
||||
duration = self.frames_written / f.samplerate if f.samplerate else 0.0
|
||||
return {
|
||||
"path": self.path,
|
||||
"frames": self.frames_written,
|
||||
"duration": duration,
|
||||
"bytes": self.bytes_written,
|
||||
"peak": self.peak,
|
||||
"peak_dbfs": 20 * np.log10(self.peak) if self.peak > 0 else float("-inf"),
|
||||
"clipped_samples": self.clipped_samples,
|
||||
"format": f.describe(),
|
||||
"dither": self._dither is not None,
|
||||
"rf64": self.rf64,
|
||||
}
|
||||
|
||||
def abort(self) -> None:
|
||||
"""放弃写入：关闭句柄但不回写头部（用于创建后立即失败的场景）。"""
|
||||
if not self.closed:
|
||||
try:
|
||||
self._fh.close()
|
||||
finally:
|
||||
self.closed = True
|
||||
|
||||
def __enter__(self) -> "WavWriter":
|
||||
return self
|
||||
|
||||
def __exit__(self, *exc) -> None:
|
||||
if exc and exc[0] is not None:
|
||||
self.abort()
|
||||
else:
|
||||
self.close()
|
||||
|
||||
|
||||
@dataclass
|
||||
class _Header:
|
||||
bytes: bytes
|
||||
data_offset: int
|
||||
riff_overhead: int
|
||||
ds64_offset: int | None = None
|
||||
|
||||
|
||||
class _RiffOverflow(Exception):
|
||||
"""RIFF 4 GiB 上限。引擎捕获后自动分段。"""
|
||||
|
||||
|
||||
def _monotonic() -> float:
|
||||
import time
|
||||
return time.monotonic()
|
||||
|
||||
|
||||
class WavReader:
|
||||
"""轻量 WAV 读取器，支持 PCM 16/24/32、float32、EXTENSIBLE、RF64。"""
|
||||
|
||||
def __init__(self, path: str):
|
||||
self.path = os.fspath(path)
|
||||
self.fh = open(self.path, "rb")
|
||||
self._parse()
|
||||
|
||||
def _parse(self) -> None:
|
||||
fh = self.fh
|
||||
hdr = fh.read(12)
|
||||
if len(hdr) < 12:
|
||||
raise ValueError("不是有效的 WAV 文件（文件过短）")
|
||||
riff, _size, wave = hdr[:4], hdr[4:8], hdr[8:12]
|
||||
if riff == b"RF64":
|
||||
self.rf64 = True
|
||||
elif riff == b"RIFF":
|
||||
self.rf64 = False
|
||||
else:
|
||||
raise ValueError("不是有效的 WAV 文件（缺少 RIFF/RF64 标记）")
|
||||
if wave != b"WAVE":
|
||||
raise ValueError("不是有效的 WAV 文件（缺少 WAVE 标记）")
|
||||
|
||||
self.fmt_tag = None
|
||||
self.channels = self.samplerate = self.bits = 0
|
||||
self.encoding = "pcm"
|
||||
self.data_offset = self.data_size = 0
|
||||
self.frames = 0
|
||||
self._ds64_data_size = None
|
||||
self._ds64_frames = None
|
||||
self._fact_frames = None
|
||||
|
||||
while True:
|
||||
cid = fh.read(4)
|
||||
if len(cid) < 4:
|
||||
break
|
||||
(csize,) = struct.unpack("<I", fh.read(4))
|
||||
body_start = fh.tell()
|
||||
if cid == b"fmt ":
|
||||
body = fh.read(csize)
|
||||
(tag, ch, sr, _bps, _align, bits) = struct.unpack("<HHIIHH", body[:16])
|
||||
if tag == WAVE_FORMAT_EXTENSIBLE and csize >= 40:
|
||||
(bits,) = struct.unpack("<H", body[18:20])
|
||||
if body[24:40] == _SUBTYPE_FLOAT:
|
||||
tag = WAVE_FORMAT_IEEE_FLOAT
|
||||
elif body[24:40] == _SUBTYPE_PCM:
|
||||
tag = WAVE_FORMAT_PCM
|
||||
self.fmt_tag, self.channels, self.samplerate = tag, ch, sr
|
||||
self.bits = bits
|
||||
self.encoding = "float" if tag == WAVE_FORMAT_IEEE_FLOAT else "pcm"
|
||||
elif cid == b"data":
|
||||
self.data_offset = body_start
|
||||
self.data_size = csize
|
||||
elif cid == b"fact" and csize >= 4:
|
||||
(self._fact_frames,) = struct.unpack("<I", fh.read(4))
|
||||
elif cid == b"ds64":
|
||||
body = fh.read(csize)
|
||||
_riff64, data64, frames64, _table = struct.unpack("<QQQI", body[:28])
|
||||
self._ds64_data_size = data64
|
||||
self._ds64_frames = frames64
|
||||
fh.seek(body_start + csize + (csize & 1))
|
||||
|
||||
if self.fmt_tag is None:
|
||||
raise ValueError("WAV 文件缺少 fmt 块")
|
||||
if self.encoding == "pcm" and self.bits not in (16, 24, 32):
|
||||
raise ValueError(f"暂不支持的 PCM 位深: {self.bits}")
|
||||
if self.rf64 and self._ds64_data_size is not None:
|
||||
self.data_size = int(self._ds64_data_size)
|
||||
if self.channels <= 0 or self.samplerate <= 0:
|
||||
raise ValueError("WAV 头部信息不完整")
|
||||
self.frames = self.data_size // (self.channels * (self.bits // 8))
|
||||
|
||||
@property
|
||||
def fmt(self) -> WavFormat:
|
||||
return WavFormat(self.samplerate, self.channels, self.encoding, self.bits)
|
||||
|
||||
@property
|
||||
def duration(self) -> float:
|
||||
return self.frames / self.samplerate if self.samplerate else 0.0
|
||||
|
||||
def read(self, start_frame: int = 0, num_frames: int | None = None) -> np.ndarray:
|
||||
"""读取为 float64 数组，形状 (frames, channels)，范围 [-1, 1]。"""
|
||||
bps = self.bits // 8
|
||||
if num_frames is None:
|
||||
num_frames = max(0, self.frames - start_frame)
|
||||
num_frames = min(num_frames, max(0, self.frames - start_frame))
|
||||
if num_frames <= 0:
|
||||
return np.zeros((0, self.channels), dtype=np.float64)
|
||||
self.fh.seek(self.data_offset + start_frame * self.channels * bps)
|
||||
raw = self.fh.read(num_frames * self.channels * bps)
|
||||
flat = pcm_to_float(raw, self.encoding, self.bits)
|
||||
usable = (flat.size // self.channels) * self.channels
|
||||
return flat[:usable].reshape(-1, self.channels)
|
||||
|
||||
def iter_blocks(self, block_frames: int = 1 << 16) -> Iterator[np.ndarray]:
|
||||
"""分块迭代读取，便于处理超长录音而不占内存。"""
|
||||
pos = 0
|
||||
while pos < self.frames:
|
||||
block = self.read(pos, block_frames)
|
||||
if block.size == 0:
|
||||
break
|
||||
yield block
|
||||
pos += block.shape[0]
|
||||
|
||||
def close(self) -> None:
|
||||
if not self.fh.closed:
|
||||
self.fh.close()
|
||||
|
||||
def __enter__(self) -> "WavReader":
|
||||
return self
|
||||
|
||||
def __exit__(self, *exc) -> None:
|
||||
self.close()
|
||||
|
||||
def info(self) -> dict:
|
||||
return {
|
||||
"path": self.path,
|
||||
"samplerate": self.samplerate,
|
||||
"channels": self.channels,
|
||||
"encoding": self.encoding,
|
||||
"bits": self.bits,
|
||||
"frames": self.frames,
|
||||
"duration": self.duration,
|
||||
"data_bytes": self.data_size,
|
||||
"rf64": self.rf64,
|
||||
"format": self.fmt.describe(),
|
||||
}
|
||||
|
||||
|
||||
def wav_info(path: str) -> dict:
|
||||
with WavReader(path) as r:
|
||||
return r.info()
|
||||
|
||||
|
||||
def read_wav(path: str, *, max_seconds: float | None = None) -> tuple[np.ndarray, int]:
|
||||
"""读取整个 WAV，返回 ``(data, samplerate)``；``max_seconds`` 可只读前段。"""
|
||||
with WavReader(path) as r:
|
||||
n = None if max_seconds is None else int(max_seconds * r.samplerate)
|
||||
return r.read(0, n), r.samplerate
|
||||
|
||||
|
||||
def write_wav(path: str, data: np.ndarray, samplerate: int, *,
|
||||
bit_depth: int | str = 24, dither: bool = True,
|
||||
rf64: bool = False) -> dict:
|
||||
"""一次性写出 WAV（离线后期处理用）。"""
|
||||
x = np.asarray(data, dtype=np.float32)
|
||||
if x.ndim == 1:
|
||||
x = x[:, None]
|
||||
encoding, bits = _fmt_for_bitdepth(bit_depth)
|
||||
fmt = WavFormat(int(samplerate), int(x.shape[1]), encoding, bits)
|
||||
w = WavWriter(path, fmt, dither=dither, rf64=rf64)
|
||||
try:
|
||||
# 分块写，避免超大数组一次性编码造成内存峰值
|
||||
step = 1 << 18
|
||||
for i in range(0, x.shape[0], step):
|
||||
w.write(x[i:i + step])
|
||||
finally:
|
||||
stats = w.close()
|
||||
stats["format"] = fmt.describe()
|
||||
return stats
|
||||
Reference in New Issue
Block a user