Make whisper an optional extra with faster-whisper by default

[Dreamsorcerer]

Problem

Whisper requires downloading a 150MB model and depends on torch (with GBs of CUDA downloads).

Solution

Provide faster-whisper by default (2MB) and use as a fallback when whisper is not available.
This avoids the 150MB download, and means we are one step closer to not depending on torch for a base install.

Breaking Changes

Users need to request dimos[whisper] now for full whisper feature.

Test

python -c "
from dimos.stream.audio.pipelines import stt
node = stt()
node.emit_text().subscribe(on_next=lambda t: print(f"USER: {t}"))
from dimos.stream.audio.utils import keepalive
keepalive()
"

[Dreamsorcerer]

TTS seems to work pretty well with faster-whisper anyway.

[greptile-apps]

Greptile Summary

This PR makes openai-whisper an optional extra (dimos[whisper]) and adds faster-whisper as the default audio transcription backend in dimos[agents], significantly reducing the default install footprint. The WhisperNode class now auto-detects which backend is available at import time, preferring openai-whisper if present and falling back to faster-whisper otherwise.

• The UserWarning on line 36–40 fires for every default install (faster-whisper is in agents), misleading users into thinking their setup is degraded when it is the intended configuration.
• faster-whisper in pyproject.toml has no lower version bound, but device=\"auto\" requires >=1.0.0, which can cause a TypeError at runtime on older installs.

Confidence Score: 4/5

Safe to merge after fixing the misleading UserWarning that fires for all default users.

One P1 finding: the UserWarning implies a degraded fallback state for every default install, which will confuse users. The missing version constraint on faster-whisper (P2) could also cause a runtime TypeError with older versions. Both are straightforward one-line fixes.

dimos/stream/audio/stt/node_whisper.py (misleading warning), pyproject.toml (missing version constraint)

Important Files Changed

Filename	Overview
dimos/stream/audio/stt/node_whisper.py	Adds faster-whisper fallback when openai-whisper is absent; misleading UserWarning fires for all default installs, and caller's modelopts dict is mutated via pop().
pyproject.toml	Moves openai-whisper to a new optional [whisper] extra and adds faster-whisper to [agents]; no version constraint on faster-whisper despite using device="auto" (requires >=1.0.0).
uv.lock	Lockfile updated to reflect new faster-whisper dependency; no manual review needed.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[import WhisperNode] --> B{try: import whisper}
    B -- success --> C[_USE_FASTER_WHISPER = False\nopenai-whisper backend]
    B -- ImportError --> D{try: from faster_whisper\nimport WhisperModel}
    D -- success --> E[UserWarning fired\n_USE_FASTER_WHISPER = True\nfaster-whisper backend]
    D -- ImportError --> F[Raise ImportError\nNo backend found]

    C --> G[WhisperNode.__init__]
    E --> G
    G --> H{_USE_FASTER_WHISPER?}
    H -- True --> I[pop fp16 → compute_type\nWhisperModel device=auto]
    H -- False --> J[whisper.load_model]

    I --> K[transcribe → segments iterator\njoin seg.text]
    J --> L[transcribe → dict\nresult text]

Loading

_{Reviews (1): Last reviewed commit: "Add warning" | Re-trigger Greptile}

[greptile-apps]

Warning fires for all default users

faster-whisper is now bundled in the agents extra, meaning it is the intended default backend. Every user who installs the standard package will have only faster-whisper available, so this UserWarning fires for 100% of default users on every first import. The message "falling back to faster-whisper" implies a degraded or misconfigured setup, which will confuse users who have done nothing wrong.

Consider removing the warning entirely (faster-whisper is the intended default, not a fallback), or downgrade it to logging.debug, or reword it so it only surfaces if the user somehow expected openai-whisper:

Suggested change

	warnings.warn(
	"openai-whisper not installed, falling back to faster-whisper. "
	"Install openai-whisper for the full backend: pip install dimos[whisper]",
	stacklevel=2,
	)
	_USE_FASTER_WHISPER = True
	warnings.warn(
	"openai-whisper is not installed; using faster-whisper. "
	"Install openai-whisper via `pip install dimos[whisper]` to use it instead.",
	stacklevel=2,
	)

Or simply omit the warning since this is the documented default path.

[greptile-apps]

Caller's dict mutated by .pop()

modelopts.pop("fp16", False) mutates the dict that was passed in. If the caller stored a reference to their options dict and inspects it after construction, the fp16 key will be silently gone. A shallow copy before popping avoids this:

Suggested change

	if _USE_FASTER_WHISPER:
	fp16 = modelopts.pop("fp16", False)
	compute_type = "float16" if fp16 else "int8"
	self.modelopts = modelopts
	self.model = WhisperModel(model, device="auto", compute_type=compute_type)
	opts = dict(modelopts)
	fp16 = opts.pop("fp16", False)
	compute_type = "float16" if fp16 else "int8"
	self.modelopts = opts

[greptile-apps]

No version constraint on faster-whisper

faster-whisper is listed without a minimum version. The device="auto" argument passed to WhisperModel (line 71 of node_whisper.py) was introduced in faster-whisper>=1.0.0. Without a lower bound, users on older versions will get a runtime TypeError. Adding a floor constraint prevents this:

Suggested change

	"faster-whisper",
	"faster-whisper>=1.0.0",

[paul-nechifor]

Why include dimos[agents] here?

[Dreamsorcerer]

Does it make sense to install without agents? I was just assuming that we'd want all the other dependencies here.