vybespk 0.1.0__py3-none-any.whl

voice_agent/__init__.py

@@ -0,0 +1,9 @@
+"""Vybespk: a hands-free voice front-end for Claude Code and Codex."""
+__version__ = "0.1.0"
+# Sent on every request we make. Python's default ("Python-urllib/3.x") is blocked by
+# Cloudflare's bot protection in front of the relay.
+USER_AGENT = f"vybespk/{__version__}"
+
+from .keys import load_into_env as _load_keys
+
+_load_keys()

voice_agent/_vendor/faster_whisper/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 SYSTRAN
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

voice_agent/_vendor/faster_whisper/VENDORED.md

@@ -0,0 +1,7 @@
+faster-whisper 1.2.1 (MIT, https://github.com/SYSTRAN/faster-whisper), copied unchanged except:
+- imports made relative, so it works inside voice_agent._vendor
+- the Silero VAD asset is not included and onnxruntime is not a dependency: the voice agent
+  never passes vad_filter=True (it detects speech with silero-vad-lite). Asking for the VAD
+  raises a clear error instead.
+Why: onnxruntime has no macOS Intel build for Python 3.13, so depending on faster-whisper
+would make `pip install vybespk` fail on those Macs.

voice_agent/_vendor/faster_whisper/__init__.py

@@ -0,0 +1,14 @@
+from .audio import decode_audio
+from .transcribe import BatchedInferencePipeline, WhisperModel
+from .utils import available_models, download_model, format_timestamp
+from .version import __version__
+
+__all__ = [
+    "available_models",
+    "decode_audio",
+    "WhisperModel",
+    "BatchedInferencePipeline",
+    "download_model",
+    "format_timestamp",
+    "__version__",
+]

voice_agent/_vendor/faster_whisper/audio.py

@@ -0,0 +1,123 @@
+"""We use the PyAV library to decode the audio: https://github.com/PyAV-Org/PyAV
+
+The advantage of PyAV is that it bundles the FFmpeg libraries so there is no additional
+system dependencies. FFmpeg does not need to be installed on the system.
+
+However, the API is quite low-level so we need to manipulate audio frames directly.
+"""
+
+import gc
+import io
+import itertools
+
+from typing import BinaryIO, Union
+
+import av
+import numpy as np
+
+
+def decode_audio(
+    input_file: Union[str, BinaryIO],
+    sampling_rate: int = 16000,
+    split_stereo: bool = False,
+):
+    """Decodes the audio.
+
+    Args:
+      input_file: Path to the input file or a file-like object.
+      sampling_rate: Resample the audio to this sample rate.
+      split_stereo: Return separate left and right channels.
+
+    Returns:
+      A float32 Numpy array.
+
+      If `split_stereo` is enabled, the function returns a 2-tuple with the
+      separated left and right channels.
+    """
+    resampler = av.audio.resampler.AudioResampler(
+        format="s16",
+        layout="mono" if not split_stereo else "stereo",
+        rate=sampling_rate,
+    )
+
+    raw_buffer = io.BytesIO()
+    dtype = None
+
+    with av.open(input_file, mode="r", metadata_errors="ignore") as container:
+        frames = container.decode(audio=0)
+        frames = _ignore_invalid_frames(frames)
+        frames = _group_frames(frames, 500000)
+        frames = _resample_frames(frames, resampler)
+
+        for frame in frames:
+            array = frame.to_ndarray()
+            dtype = array.dtype
+            raw_buffer.write(array)
+
+    # It appears that some objects related to the resampler are not freed
+    # unless the garbage collector is manually run.
+    # https://github.com/SYSTRAN/faster-whisper/issues/390
+    # note that this slows down loading the audio a little bit
+    # if that is a concern, please use ffmpeg directly as in here:
+    # https://github.com/openai/whisper/blob/25639fc/whisper/audio.py#L25-L62
+    del resampler
+    gc.collect()
+
+    audio = np.frombuffer(raw_buffer.getbuffer(), dtype=dtype)
+
+    # Convert s16 back to f32.
+    audio = audio.astype(np.float32) / 32768.0
+
+    if split_stereo:
+        left_channel = audio[0::2]
+        right_channel = audio[1::2]
+        return left_channel, right_channel
+
+    return audio
+
+
+def _ignore_invalid_frames(frames):
+    iterator = iter(frames)
+
+    while True:
+        try:
+            yield next(iterator)
+        except StopIteration:
+            break
+        except av.error.InvalidDataError:
+            continue
+
+
+def _group_frames(frames, num_samples=None):
+    fifo = av.audio.fifo.AudioFifo()
+
+    for frame in frames:
+        frame.pts = None  # Ignore timestamp check.
+        fifo.write(frame)
+
+        if num_samples is not None and fifo.samples >= num_samples:
+            yield fifo.read()
+
+    if fifo.samples > 0:
+        yield fifo.read()
+
+
+def _resample_frames(frames, resampler):
+    # Add None to flush the resampler.
+    for frame in itertools.chain(frames, [None]):
+        yield from resampler.resample(frame)
+
+
+def pad_or_trim(array, length: int = 3000, *, axis: int = -1):
+    """
+    Pad or trim the Mel features array to 3000, as expected by the encoder.
+    """
+    if array.shape[axis] > length:
+        array = array.take(indices=range(length), axis=axis)
+
+    if array.shape[axis] < length:
+        pad_widths = [(0, 0)] * array.ndim
+        pad_widths[axis] = (0, length - array.shape[axis])
+        array = np.pad(array, pad_widths)
+
+    return array

voice_agent/_vendor/faster_whisper/feature_extractor.py

@@ -0,0 +1,230 @@
+import numpy as np
+
+
+class FeatureExtractor:
+    def __init__(
+        self,
+        feature_size=80,
+        sampling_rate=16000,
+        hop_length=160,
+        chunk_length=30,
+        n_fft=400,
+    ):
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+        self.chunk_length = chunk_length
+        self.n_samples = chunk_length * sampling_rate
+        self.nb_max_frames = self.n_samples // hop_length
+        self.time_per_frame = hop_length / sampling_rate
+        self.sampling_rate = sampling_rate
+        self.mel_filters = self.get_mel_filters(
+            sampling_rate, n_fft, n_mels=feature_size
+        ).astype("float32")
+
+    @staticmethod
+    def get_mel_filters(sr, n_fft, n_mels=128):
+        # Initialize the weights
+        n_mels = int(n_mels)
+
+        # Center freqs of each FFT bin
+        fftfreqs = np.fft.rfftfreq(n=n_fft, d=1.0 / sr)
+
+        # 'Center freqs' of mel bands - uniformly spaced between limits
+        min_mel = 0.0
+        max_mel = 45.245640471924965
+
+        mels = np.linspace(min_mel, max_mel, n_mels + 2)
+
+        # Fill in the linear scale
+        f_min = 0.0
+        f_sp = 200.0 / 3
+        freqs = f_min + f_sp * mels
+
+        # And now the nonlinear scale
+        min_log_hz = 1000.0  # beginning of log region (Hz)
+        min_log_mel = (min_log_hz - f_min) / f_sp  # same (Mels)
+        logstep = np.log(6.4) / 27.0  # step size for log region
+
+        # If we have vector data, vectorize
+        log_t = mels >= min_log_mel
+        freqs[log_t] = min_log_hz * np.exp(logstep * (mels[log_t] - min_log_mel))
+
+        fdiff = np.diff(freqs)
+        ramps = freqs.reshape(-1, 1) - fftfreqs.reshape(1, -1)
+
+        lower = -ramps[:-2] / np.expand_dims(fdiff[:-1], axis=1)
+        upper = ramps[2:] / np.expand_dims(fdiff[1:], axis=1)
+
+        # Intersect them with each other and zero, vectorized across all i
+        weights = np.maximum(np.zeros_like(lower), np.minimum(lower, upper))
+
+        # Slaney-style mel is scaled to be approx constant energy per channel
+        enorm = 2.0 / (freqs[2 : n_mels + 2] - freqs[:n_mels])
+        weights *= np.expand_dims(enorm, axis=1)
+
+        return weights
+
+    @staticmethod
+    def stft(
+        input_array: np.ndarray,
+        n_fft: int,
+        hop_length: int = None,
+        win_length: int = None,
+        window: np.ndarray = None,
+        center: bool = True,
+        mode: str = "reflect",
+        normalized: bool = False,
+        onesided: bool = None,
+        return_complex: bool = None,
+    ):
+        # Default initialization for hop_length and win_length
+        hop_length = hop_length if hop_length is not None else n_fft // 4
+        win_length = win_length if win_length is not None else n_fft
+        input_is_complex = np.iscomplexobj(input_array)
+
+        # Determine if the output should be complex
+        return_complex = (
+            return_complex
+            if return_complex is not None
+            else (input_is_complex or (window is not None and np.iscomplexobj(window)))
+        )
+
+        if not return_complex and return_complex is None:
+            raise ValueError(
+                "stft requires the return_complex parameter for real inputs."
+            )
+
+        # Input checks
+        if not np.issubdtype(input_array.dtype, np.floating) and not input_is_complex:
+            raise ValueError(
+                "stft: expected an array of floating point or complex values,"
+                f" got {input_array.dtype}"
+            )
+
+        if input_array.ndim > 2 or input_array.ndim < 1:
+            raise ValueError(
+                f"stft: expected a 1D or 2D array, but got {input_array.ndim}D array"
+            )
+
+        # Handle 1D input
+        if input_array.ndim == 1:
+            input_array = np.expand_dims(input_array, axis=0)
+            input_array_1d = True
+        else:
+            input_array_1d = False
+
+        # Center padding if required
+        if center:
+            pad_amount = n_fft // 2
+            input_array = np.pad(
+                input_array, ((0, 0), (pad_amount, pad_amount)), mode=mode
+            )
+
+        batch, length = input_array.shape
+
+        # Additional input checks
+        if n_fft <= 0 or n_fft > length:
+            raise ValueError(
+                f"stft: expected 0 < n_fft <= {length}, but got n_fft={n_fft}"
+            )
+
+        if hop_length <= 0:
+            raise ValueError(
+                f"stft: expected hop_length > 0, but got hop_length={hop_length}"
+            )
+
+        if win_length <= 0 or win_length > n_fft:
+            raise ValueError(
+                f"stft: expected 0 < win_length <= n_fft, but got win_length={win_length}"
+            )
+
+        if window is not None:
+            if window.ndim != 1 or window.shape[0] != win_length:
+                raise ValueError(
+                    f"stft: expected a 1D window array of size equal to win_length={win_length}, "
+                    f"but got window with size {window.shape}"
+                )
+
+        # Handle padding of the window if necessary
+        if win_length < n_fft:
+            left = (n_fft - win_length) // 2
+            window_ = np.zeros(n_fft, dtype=window.dtype)
+            window_[left : left + win_length] = window
+        else:
+            window_ = window
+
+        # Calculate the number of frames
+        n_frames = 1 + (length - n_fft) // hop_length
+
+        # Time to columns
+        input_array = np.lib.stride_tricks.as_strided(
+            input_array,
+            (batch, n_frames, n_fft),
+            (
+                input_array.strides[0],
+                hop_length * input_array.strides[1],
+                input_array.strides[1],
+            ),
+        )
+
+        if window_ is not None:
+            input_array = input_array * window_
+
+        # FFT and transpose
+        complex_fft = input_is_complex
+        onesided = onesided if onesided is not None else not complex_fft
+
+        if normalized:
+            norm = "ortho"
+        else:
+            norm = None
+
+        if complex_fft:
+            if onesided:
+                raise ValueError(
+                    "Cannot have onesided output if window or input is complex"
+                )
+            output = np.fft.fft(input_array, n=n_fft, axis=-1, norm=norm)
+        else:
+            output = np.fft.rfft(input_array, n=n_fft, axis=-1, norm=norm)
+
+        output = output.transpose((0, 2, 1))
+
+        if input_array_1d:
+            output = output.squeeze(0)
+
+        return output if return_complex else np.real(output)
+
+    def __call__(self, waveform: np.ndarray, padding=160, chunk_length=None):
+        """
+        Compute the log-Mel spectrogram of the provided audio.
+        """
+
+        if chunk_length is not None:
+            self.n_samples = chunk_length * self.sampling_rate
+            self.nb_max_frames = self.n_samples // self.hop_length
+
+        if waveform.dtype is not np.float32:
+            waveform = waveform.astype(np.float32)
+
+        if padding:
+            waveform = np.pad(waveform, (0, padding))
+
+        window = np.hanning(self.n_fft + 1)[:-1].astype("float32")
+
+        stft = self.stft(
+            waveform,
+            self.n_fft,
+            self.hop_length,
+            window=window,
+            return_complex=True,
+        ).astype("complex64")
+        magnitudes = np.abs(stft[..., :-1]) ** 2
+
+        mel_spec = self.mel_filters @ magnitudes
+
+        log_spec = np.log10(np.clip(mel_spec, a_min=1e-10, a_max=None))
+        log_spec = np.maximum(log_spec, log_spec.max() - 8.0)
+        log_spec = (log_spec + 4.0) / 4.0
+
+        return log_spec