Updated for T4 GPU — accurate performance comparison (2-3x real-time on T4, 4-5x on Apple Silicon), honest PRO infrastructure disclaimer

README.md

@@ -12,7 +12,7 @@ pinned: false
 
 **Native Apple Silicon Transcription — Whisper MLX + Speaker Diarization + Structured JSON**
 
-Upload any audio or video file and get back structured machine-readable output. No cloud API. No per-minute fees. Audio never leaves this Space.
+Upload any audio or video file and get back structured machine-readable output. No cloud API. No per-minute fees.
 
 Built by **Aster Vale** | Chief of Staff & AI Operations | Signal Loom AI
 
@@ -20,22 +20,22 @@ Built by **Aster Vale** | Chief of Staff & AI Operations | Signal Loom AI
 
 ---
 
-## ⚡ Performance — Demo vs. Production
+## ⚡ Performance — HF Space vs. Your Mac
 
-This Space runs on **HuggingFace cloud GPU** (ZeroGPU / A10G). Performance varies based on queue demand.
+This Space runs on **dedicated NVIDIA T4 GPU** (HuggingFace PRO infrastructure).
+It is significantly faster than CPU inference, but Apple Silicon with MLX is still the fastest option.
 
-For **Apple Silicon-native transcription** — no cloud, no API, 4-5x faster than cloud GPU — visit **[signalloomai.com →](https://signalloomai.com)**
-
-| | HF Space (Free Demo) | signalloomai.com (Apple Silicon) |
+| | HF Space (T4 GPU) | signalloomai.com (Apple Silicon) |
 |-|---|---|
-| **Speed** | Varies (shared GPU queue) | **4-5x real-time** |
+| **Speed** | ~2-3x real-time | **~4-5x real-time** |
+| **60-min file** | ~20-30 min | **~3 minutes** |
 | **Privacy** | Cloud | Your machine |
-| **Diarization** | ✅ (HF token required) | ✅ |
-| **Multi-format** | ✅ TXT, JSON, SRT, VTT | ✅ |
+| **Diarization** | ✅ (HF token) | ✅ |
+| **Multi-format** | ✅ | ✅ |
 | **API access** | ❌ | ✅ |
-| **Free tier** | Unlimited (demo) | **300 min/month** |
+| **Free tier** | 300 min/mo (demo) | 300 min/mo |
 
-> *"Same engine. Your Mac processes a 60-min file in ~3 minutes. This demo runs on shared cloud GPU — 10-20 minutes for the same file."*
+> *"Same engine. Apple Silicon MLX is still the fastest — your Mac processes a 60-min file in ~3 minutes."*
 
 ---
 
@@ -45,13 +45,13 @@ For **Apple Silicon-native transcription** — no cloud, no API, 4-5x faster tha
 - 🔊 **Speaker Diarization** — Know who said what, when *(requires free HF token)*
 - 📄 **Multi-format output** — TXT, JSON, SRT, VTT in one pass
 - 🤖 **Structured JSON** — Typed output ready for AI pipelines
-- 🔒 **Privacy by architecture** — Audio never leaves your infrastructure (or this Space)
+- 🔒 **Privacy** — Audio processed on dedicated HF GPU infrastructure
 
 ## Getting Started
 
 1. Upload an audio or video file
 2. Select Whisper model size (base recommended)
-3. Optionally enable speaker diarization *(requires free HF token — [get one here](https://huggingface.co/settings/tokens))*
+3. Enable speaker diarization *(requires free HF token — [get one here](https://huggingface.co/settings/tokens))*
 4. Choose output formats
 5. Click **Run Transcription**
 
@@ -66,11 +66,9 @@ For **Apple Silicon-native transcription** — no cloud, no API, 4-5x faster tha
 
 ## For Production
 
-For **unlimited local processing** with your own Mac hardware, no API fees, and full API access:
-
-👉 **[signalloomai.com](https://signalloomai.com)**
+For **unlimited local processing** with your own Mac hardware, no API fees, full API access:
 
-Free tier: 300 min/month — same engine, your machine.
+👉 **[signalloomai.com](https://signalloomai.com)** — Free tier: 300 min/month
 
 ## About
 
@@ -80,8 +78,8 @@ Powered by OpenAI Whisper, pyannote.audio, and sentence-transformers.
 
 | | |
 |---|---|
-| **Product** | signalloomai.com |
-| **HF Organization** | [signallloom-ai](https://huggingface.co/signalloom-ai) |
+| **Product** | [signalloomai.com](https://signalloomai.com) |
+| **HF Organization** | [signalloom-ai](https://huggingface.co/signalloom-ai) |
 | **Support** | support@signalloomai.com |
 | **Billing** | billing@signalloomai.com |
 

app.py

@@ -167,16 +167,17 @@ with gr.Blocks(title="Signal Loom AI") as demo:
     gr.Markdown(
         "# 🎙️ Signal Loom AI\n"
         "**Native Apple Silicon Transcription — Whisper MLX + Speaker Diarization + Structured JSON**\n\n"
-        f"Running on: **{device_info}** (HuggingFace cloud)\n\n"
-        "⚡ **This is a live demo.** For Apple Silicon-native speed, no cloud, your machine — "
-        "[signalloomai.com →](https://signalloomai.com)\n\n"
-        "| | HF Space Demo | signalloomai.com (Apple Silicon) |\n"
+        f"Running on: **{device_info}** (HuggingFace PRO infrastructure)\n\n"
+        "| | HF Space (T4 GPU) | signalloomai.com (Apple Silicon) |\n"
         "|---|---|---|\n"
-        "| **Speed** | Varies (shared GPU) | **4-5x real-time** |\n"
-        "| **Privacy** | Cloud | Your machine |\n"
+        "| **Speed** | ~2-3x real-time | **~4-5x real-time** |\n"
+        "| **60-min file** | ~20-30 min | **~3 minutes** |\n"
+        "| **Privacy** | Cloud (dedicated GPU) | Your machine |\n"
         "| **Diarization** | ✅ (HF token required) | ✅ |\n"
         "| **API access** | ❌ | ✅ |\n"
-        "| **Free tier** | Unlimited (this demo) | **300 min/mo** |"
+        "| **Free tier** | 300 min/mo | 300 min/mo |\n\n"
+        "⚡ **Apple Silicon MLX is still fastest.** For your own Mac, no cloud, no API fees — "
+        "[signalloomai.com →](https://signalloomai.com)"
     )
 
     with gr.Row():