chore: drop unused faster-whisper/ctranslate2/requests deps

Voice posing runs entirely in the browser via the Web Speech API
(fusion_ui.BROWSER_VOICE_JS); the server-side faster-whisper path in
src/buddy_fusion/asr.py is never imported by the app (verified: grep finds
no `import asr`, git history shows fusion_ui never imported it). The
faster-whisper / ctranslate2 / requests pins only served that dead code, so
remove them as dead weight on the Space image. requirements.txt keeps a
restore recipe in comments.

Document the finding in wiki/voice-input-asr.md (corrects an earlier draft
that wrongly claimed the Space runs faster-whisper).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

requirements.txt

@@ -13,17 +13,11 @@ llama-cpp-python==0.3.19
 gradio==6.15.2
 huggingface_hub>=1.17
 spaces
-# faster-whisper (CTranslate2) powers push-to-talk voice posing. Runs on
-# ZeroGPU inside @spaces.GPU on the Space; CPU/int8 locally.
-# faster-whisper's code imports `requests`, but it only declares
-# huggingface-hub>=0.13; hf_hub 1.x (pinned above) switched to httpx and no
-# longer pulls requests transitively, so add it explicitly or the ASR import
-# fails on the Space.
-faster-whisper==1.1.0
-# CTranslate2 is faster-whisper's CUDA backend (a non-PyTorch CUDA lib, same
-# class of pin hazard as llama-cpp-python). >=4.5 needs cuDNN 9 + CUDA>=12.3,
-# which torch==2.8.0 (cu128) provides. 4.8.0 is verified importable locally;
-# confirm on the Space build log it installs without a cuDNN/CUDA error. If the
-# Space build fails on cuDNN 9, 4.4.0 (cuDNN 8) is the fallback.
-ctranslate2==4.8.0
-requests
+# NOTE: voice posing runs entirely in the browser (Web Speech API). The
+# server-side faster-whisper path in src/buddy_fusion/asr.py is never imported
+# by the app, so faster-whisper / ctranslate2 (its CUDA backend) and the
+# requests pin (added only so the unused ASR import would resolve) were removed
+# as dead weight. See wiki/voice-input-asr.md. To restore the Whisper path:
+#   faster-whisper==1.1.0
+#   ctranslate2==4.8.0   # >=4.5 needs cuDNN 9 + CUDA>=12.3 (torch 2.8.0 cu128); 4.4.0 = cuDNN 8 fallback
+#   requests             # hf_hub 1.x uses httpx and no longer pulls requests transitively

wiki/index.md

@@ -26,6 +26,9 @@ See [README](README.md) for how this wiki works.
   and the CPU-vs-ZeroGPU baseline.
 - [gradio-embed-integration](gradio-embed-integration.md) - The Three.js voxel
   embed init and the JS↔Python voice bridge inside `gr.HTML`.
+- [voice-input-asr](voice-input-asr.md) - Voice runs on browser Web Speech, NOT
+  faster-whisper; asr.py is unwired dead code and the faster-whisper dependency
+  is never exercised.
 
 ## Debugging
 

wiki/log.md

@@ -2,6 +2,18 @@
 
 ## 2026-06-14
 
+### Voice input ASR page (corrected)
+
+- Added `voice-input-asr.md`. Initial draft wrongly claimed the Space runs
+  faster-whisper (inferred from `asr.py` defaults + `requirements.txt`).
+  Strict verification corrected this: `grep -rn "import asr"` returns zero
+  matches and `git log -S "import asr"` shows `fusion_ui` never imported it —
+  `asr.py` is unwired dead code. The live voice path is browser Web Speech
+  (`window.SpeechRecognition` in `fusion_ui.BROWSER_VOICE_JS`); the app only
+  ever handles the plain-text transcript. `faster-whisper`/`ctranslate2` are
+  installed but never exercised; `BUDDY_ASR_*` env vars have no effect.
+- Linked it into `index.md` Integrations.
+
 ### Voxel rendering page + turntable centering fix
 
 - Added `voxel-rendering.md` documenting the voxelizer, bounds-driven camera,

wiki/voice-input-asr.md

@@ -0,0 +1,72 @@
+---
+title: Voice Input — Browser Web Speech (not faster-whisper)
+tags: [architecture, asr, voice, deployment]
+last_updated: 2026-06-14
+confidence: high
+sources:
+  - src/buddy_fusion/fusion_ui.py
+  - src/buddy_fusion/asr.py
+  - requirements.txt
+related_pages:
+  - gradio-embed-integration.md
+  - system-overview.md
+  - deployment-strategy.md
+---
+
+# Voice Input — Browser Web Speech (not faster-whisper)
+
+## Summary
+
+Voice posing runs **entirely in the user's browser** via the Web Speech API.
+The Space does **not** run faster-whisper for voice, despite `faster-whisper`
+being in `requirements.txt` and a full server-side `asr.py` existing — that
+module is **never imported by the app**. This page exists because the dependency
+list and `asr.py` strongly imply a Whisper path that the wiring does not back.
+
+## What actually runs
+
+The only live voice path is the browser bridge in
+[fusion_ui.py](../src/buddy_fusion/fusion_ui.py):
+
+1. `BROWSER_VOICE_JS` (injected through `gr.HTML` head) starts
+   `window.SpeechRecognition || window.webkitSpeechRecognition` — browser-native
+   recognition, client-side, **no audio is uploaded to the server**.
+2. A final transcript is written into a hidden Gradio textbox
+   (`#buddy-voice-transcript-in`) via the native value setter, then a hidden
+   submit button (`#buddy-voice-submit`) is clicked.
+3. `voice_submit.click(on_voice, ...)` fires; `on_voice` receives the **plain
+   text** transcript and passes it to `runtime.interpret_genome_command`. No
+   `gr.Audio` component, no `transcribe()` call, no server-side STT.
+
+So on the Space, speech-to-text is performed by the user's browser (Chrome et
+al.), not by the app. Recognition quality and language coverage come from the
+browser, not from any model the Space hosts.
+
+## Why asr.py / faster-whisper are misleading
+
+- `asr.py` wraps faster-whisper (`large-v3-turbo`, `@spaces.GPU`) behind a
+  `transcribe()` surface and reads `BUDDY_ASR_BACKEND`. It is a complete,
+  working implementation — but **dead code**: `grep -rn "import asr"` across
+  src + tests + `app.py` returns zero matches, and `git log -S "import asr"`
+  shows `fusion_ui` never imported it. It has shipped unwired since the initial
+  commit (357e825), most likely ported from build-a-buddy-hf's race_ui, which
+  did use a server-side path.
+- `requirements.txt` still pins `faster-whisper==1.1.0` + `ctranslate2==4.8.0`.
+  These are installed on the Space but never exercised by any live code path —
+  dead weight on the image.
+- `BUDDY_ASR_BACKEND` / `BUDDY_ASR_MODEL` / `BUDDY_ASR_BROWSER_LANG` are read
+  only inside `asr.py`, so setting them has **no effect** on the running app.
+
+## Verification note
+
+Do not infer the runtime ASR backend from `asr.py` defaults or
+`requirements.txt`. Confirm wiring instead: search for an `import asr` / `from
+.asr` and for a `gr.Audio` component feeding `transcribe()`. As of 2026-06-14
+neither exists; the live path is browser Web Speech only.
+
+## If the Whisper path is ever wanted
+
+Wiring `asr.py` back in would need: a `gr.Audio(sources=["microphone"])`
+component, an event that calls `asr.transcribe_with_language` on the recorded
+file, and a `backend()` check to choose browser vs server. Until then, treat the
+faster-whisper dependency as removable if image size matters.