cade-cli 0.15.2__tar.gz → 0.15.5__tar.gz

{cade_cli-0.15.2 → cade_cli-0.15.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: cade-cli
-Version: 0.15.2
+Version: 0.15.5
 Summary: Cade - The CLI Agent from Arcade.dev
 Project-URL: Homepage, https://arcade.dev
 Project-URL: Documentation, https://docs.arcade.dev
@@ -123,7 +123,7 @@ The install script detects your OS/arch, downloads the matching prebuilt binary
 `~/.local/bin` — no GitHub token required. Pin a version or install location:
 
 ```bash
-CADE_VERSION=0.15.2 CADE_INSTALL_DIR=/usr/local/bin \
+CADE_VERSION=0.15.5 CADE_INSTALL_DIR=/usr/local/bin \
   sh -c "$(curl -fsSL https://arcadeagent.dev/install)"
 ```
 

{cade_cli-0.15.2 → cade_cli-0.15.5}/README.md

@@ -35,7 +35,7 @@ The install script detects your OS/arch, downloads the matching prebuilt binary
 `~/.local/bin` — no GitHub token required. Pin a version or install location:
 
 ```bash
-CADE_VERSION=0.15.2 CADE_INSTALL_DIR=/usr/local/bin \
+CADE_VERSION=0.15.5 CADE_INSTALL_DIR=/usr/local/bin \
   sh -c "$(curl -fsSL https://arcadeagent.dev/install)"
 ```
 

{cade_cli-0.15.2 → cade_cli-0.15.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "cade-cli"
-version = "0.15.2"
+version = "0.15.5"
 description = "Cade - The CLI Agent from Arcade.dev"
 readme = "README.md"
 requires-python = ">=3.11"
@@ -75,6 +75,11 @@ voice = [
     "sounddevice>=0.5.5,<1.0.0",
     "numpy>=1.24.0",
     "webrtcvad>=2.0.10",
+    # NOTE: Kokoro TTS also needs spaCy's en_core_web_sm model, but it is NOT
+    # listed here — a direct-URL dependency in package metadata is rejected by
+    # PyPI on upload. The macOS release build installs it explicitly before
+    # PyInstaller (see .github/workflows/release.yml), and cade.spec bundles it
+    # so misaki never downloads it at runtime in the frozen app.
 ]
 training = [
     "torch>=2.1.0",

{cade_cli-0.15.2 → cade_cli-0.15.5}/src/cadecoder/voice/stt.py

@@ -7,8 +7,6 @@ Models are downloaded from HuggingFace on first use and cached automatically.
 import asyncio
 import logging
 import os
-import tempfile
-import wave
 
 import numpy as np
 
@@ -17,6 +15,10 @@ from cadecoder.core.errors import VoiceError
 from cadecoder.core.logging import log
 from cadecoder.voice.audio import DEFAULT_SAMPLE_RATE
 
+# Whisper always operates at 16 kHz. When given a numpy array (rather than a
+# file path) mlx-whisper does NOT resample, so the samples must already be 16 kHz.
+_WHISPER_SAMPLE_RATE = 16_000
+
 
 def _suppress_hf_logging() -> None:
     """Suppress noisy HuggingFace Hub download/progress logging.
@@ -109,17 +111,21 @@ class WhisperSTT:
             raise VoiceError("STT model not loaded — call _ensure_loaded() first")
 
         try:
-            # mlx_whisper.transcribe accepts a file path or numpy array.
-            # Write to a temp WAV file for maximum compatibility across
-            # mlx-whisper versions.
-            with tempfile.NamedTemporaryFile(suffix=".wav", delete=True) as tmp:
-                self._write_wav(tmp.name, audio, sample_rate)
-                result = self._mlx_whisper.transcribe(  # type: ignore[union-attr]
-                    tmp.name,
-                    path_or_hf_repo=self.model,
-                    language=self.language,
-                    fp16=False,
-                )
+            # Pass the raw float32 samples directly. Writing a temp WAV and
+            # passing the *path* makes mlx-whisper shell out to `ffmpeg` to
+            # decode it (not bundled in the frozen binary, and often absent on
+            # user machines). Whisper accepts a 16 kHz mono float32 array with
+            # no external tools — the recorder already produces exactly that.
+            samples = np.asarray(audio, dtype=np.float32).flatten()
+            if sample_rate != _WHISPER_SAMPLE_RATE:
+                samples = _resample(samples, sample_rate, _WHISPER_SAMPLE_RATE)
+
+            result = self._mlx_whisper.transcribe(  # type: ignore[union-attr]
+                samples,
+                path_or_hf_repo=self.model,
+                language=self.language,
+                fp16=False,
+            )
 
             text = result.get("text", "").strip()
             log.debug(f"Transcribed {len(audio) / sample_rate:.1f}s -> '{text[:80]}'")
@@ -128,13 +134,19 @@ class WhisperSTT:
         except Exception as exc:
             raise VoiceError(f"Transcription failed: {exc}") from exc
 
-    @staticmethod
-    def _write_wav(path: str, audio: np.ndarray, sample_rate: int) -> None:
-        """Write a float32 numpy array to a 16-bit PCM WAV file."""
-        # Convert float32 [-1.0, 1.0] to int16
-        pcm = (audio * 32767).clip(-32768, 32767).astype(np.int16)
-        with wave.open(path, "wb") as wf:
-            wf.setnchannels(1)
-            wf.setsampwidth(2)  # 16-bit
-            wf.setframerate(sample_rate)
-            wf.writeframes(pcm.tobytes())
+
+def _resample(samples: np.ndarray, src_rate: int, dst_rate: int) -> np.ndarray:
+    """Linearly resample mono float32 audio to ``dst_rate`` (no external deps).
+
+    Adequate for speech fed to Whisper. Only used on the rare path where the
+    recorder's rate differs from Whisper's 16 kHz.
+    """
+    if src_rate == dst_rate or samples.size == 0:
+        return samples
+    duration = samples.shape[0] / float(src_rate)
+    n_out = int(round(duration * dst_rate))
+    if n_out <= 0:
+        return np.array([], dtype=np.float32)
+    x_old = np.linspace(0.0, duration, num=samples.shape[0], endpoint=False)
+    x_new = np.linspace(0.0, duration, num=n_out, endpoint=False)
+    return np.interp(x_new, x_old, samples).astype(np.float32)