Make whisper an optional extra with faster-whisper by default

dimos/stream/audio/stt/node_whisper.py

@@ -16,7 +16,6 @@
 from typing import Any
 
 from reactivex import Observable, create, disposable
-import whisper  # type: ignore[import-untyped]
 
 from dimos.stream.audio.base import (
     AbstractAudioConsumer,
@@ -27,10 +26,31 @@
 
 logger = setup_logger()
 
+try:
+    import whisper  # type: ignore[import-untyped]
+
+    _USE_FASTER_WHISPER = False
+except ImportError:
+    try:
+        from faster_whisper import WhisperModel  # type: ignore[import-untyped]
+
+        logger.warn(
+            "openai-whisper not installed, falling back to faster-whisper. "
+            "Install openai-whisper for the full backend: pip install openai-whisper",
+        )
+        _USE_FASTER_WHISPER = True
+    except ImportError:
+        raise ImportError(
+            "No whisper backend found. "
+            "Install faster-whisper (pip install faster-whisper) "
+            "or openai-whisper (pip install dimos[whisper])."
+        )
+
 
 class WhisperNode(AbstractAudioConsumer, AbstractTextEmitter):
     """
-    A node that transcribes audio using OpenAI's Whisper model and emits the transcribed text.
+    A node that transcribes audio using OpenAI Whisper or faster-whisper and emits
+    the transcribed text. Prefers openai-whisper if installed, falls back to faster-whisper.
     """
 
     def __init__(
@@ -41,8 +61,15 @@ def __init__(
         if modelopts is None:
             modelopts = {"language": "en", "fp16": False}
         self.audio_observable = None
-        self.modelopts = modelopts
-        self.model = whisper.load_model(model)
+
+        if _USE_FASTER_WHISPER:
+            compute_type = "float16" if modelopts.get("fp16", False) else "int8"
+            modelopts = {k: v for k, v in modelopts.items() if k != "fp16"}
+            self.modelopts = modelopts
+            self.model = WhisperModel(model, device="auto", compute_type=compute_type)
+        else:
+            self.modelopts = modelopts
+            self.model = whisper.load_model(model)
 
     def consume_audio(self, audio_observable: Observable) -> "WhisperNode":  # type: ignore[type-arg]
         """
@@ -73,8 +100,15 @@ def on_subscribe(observer, scheduler):
             # Subscribe to the audio source
             def on_audio_event(event: AudioEvent) -> None:
                 try:
-                    result = self.model.transcribe(event.data.flatten(), **self.modelopts)
-                    observer.on_next(result["text"].strip())
+                    if _USE_FASTER_WHISPER:
+                        segments, _info = self.model.transcribe(
+                            event.data.flatten(), **self.modelopts
+                        )
+                        text = " ".join(seg.text.strip() for seg in segments)
+                    else:
+                        result = self.model.transcribe(event.data.flatten(), **self.modelopts)
+                        text = result["text"].strip()
+                    observer.on_next(text)
                 except Exception as e:
                     logger.error(f"Error processing audio event: {e}")
                     observer.on_error(e)

pyproject.toml

@@ -170,8 +170,8 @@ agents = [
 
     # Audio
     "openai",
-    "openai-whisper",
     "sounddevice",
+    "faster-whisper>=1.0.0",
 ]
 
 web = [
@@ -246,6 +246,7 @@ dev = [
     "requests-mock==1.12.1",
     "terminaltexteffects==0.12.2",
     "watchdog>=3.0.0",
+    "openai-whisper",
 
     # docs
     "md-babel-py==1.1.1",
@@ -391,6 +392,7 @@ module = [
     "cyclonedds.*",
     "dimos_lcm.*",
     "etils",
+    "faster_whisper",
     "geometry_msgs.*",
     "lazy_loader",
     "mujoco",