Change app_file to main.py which patches stem rendering before building UI"

README.md

@@ -5,7 +5,7 @@ colorFrom: green
 colorTo: purple
 sdk: gradio
 sdk_version: 6.13.0
-app_file: app.py
+app_file: main.py
 pinned: true
 license: mit
 suggested_hardware: zero-a10g
@@ -16,108 +16,10 @@ tags:
   - audio
   - mixing
   - stem-separation
-  - beat-detection
+  - demucs
 short_description: AI analyzes songs, plans transitions, renders DJ sets
 ---
 
 # 🎛️ AI DJ Set Builder
 
-**Upload your songs → AI analyzes each track → finds compatible pairs → plans & renders professional DJ transitions → assembles a complete set**
-
-## How It Works
-
-### 1️⃣ Upload & Analyze
-Upload audio files. Each track is analyzed for:
-- **BPM** — spectral flux onset detection + autocorrelation
-- **Musical key** — Krumhansl-Schmuckler on CQT chroma → Camelot wheel mapping
-- **Energy profile** — RMS curve, loudness dB, spectral centroid (brightness)
-- **Structural segments** — checkerboard kernel novelty on MFCC+chroma self-similarity (Foote 2000)
-- **Cue points** — segment boundaries + 16-bar phrase energy transitions, snapped to downbeats
-
-### 2️⃣ Compatibility & Set Planning
-Multi-dimensional compatibility scoring for every track pair:
-- **BPM proximity** (35%) — 94.5% of real DJ transitions are within ±5% ([arxiv:2008.10267](https://arxiv.org/abs/2008.10267))
-- **Harmonic compatibility** (30%) — Camelot wheel distance
-- **Energy flow** (20%) — smooth energy transitions
-- **Timbral similarity** (15%) — spectral centroid proximity
-
-Tracks ordered via greedy energy-arc algorithm (warm up → peak → cool down) with configurable no-repeat constraint.
-
-### 3️⃣ 15 Professional Transition Techniques
-
-Every transition is structured as a sequence of *phases* with multi-band EQ automation, volume curves, and effects sends — mirroring how a real DJ moves faders and knobs over the course of a mix.
-
-#### Foundational
-| Type | Description | Stems? |
-|------|-------------|--------|
-| `eq_crossfade` | 3-band EQ crossfade — lows swap first (S-curve), then mids, then highs. The bread-and-butter of professional DJing. | No |
-| `long_blend` | Extended 64-beat blend with slow S-curve automation on all three EQ bands plus mid-scoop to prevent muddiness. | No |
-| `bass_swap` | Surgical bass exchange on a downbeat using isolated stems — incoming kick/bass replaces outgoing while mids+highs crossfade. | Yes |
-| `slam` | Hard cut precisely on a downbeat with a 10ms anti-click micro-fade. Maximum impact for peak-time moments. | No |
-
-#### Filter-based
-| Type | Description | Stems? |
-|------|-------------|--------|
-| `resonant_filter_sweep` | Resonant LP closes on outgoing (with rising Q) while resonant HP opens on incoming — creates the classic 'whoosh'. | No |
-| `hpf_buildup` | HPF rises exponentially on outgoing (removing kick), incoming drops in full. Mimics the tension of a live breakdown. | No |
-
-#### Effects-based
-| Type | Description | Stems? |
-|------|-------------|--------|
-| `dub_echo_out` | Pioneer DJM-style dub echo — tempo-synced delay with HP filter on the feedback path, creating thin spacey echo trails. | No |
-| `reverb_wash` | Reverb send cranks up on outgoing creating a wash, tail is progressively HP-filtered while incoming S-curves in. | No |
-| `beat_repeat_stutter` | Last beat sliced at accelerating subdivisions (1/4→1/8→1/16→1/32) creating a stutter/riser effect. | No |
-| `spinback` | Simulated vinyl spinback — rapid pitch drop to silence while incoming fades in. Short, dramatic, playful. | No |
-
-#### Stem-based (require demucs)
-| Type | Description | Stems? |
-|------|-------------|--------|
-| `acapella_over_instrumental` | Outgoing vocals isolated and layered over incoming instrumental. Creates a bridge between both tracks. | Yes |
-| `drums_first` | Incoming drums/percussion enter first, then bass swaps, then mids+highs — a layered reveal. | Yes |
-| `double_drop` | Both tracks at full energy with complementary EQ carving — one gets bass, the other gets highs. Combined super-drop. | Yes |
-
-#### Creative / tension-and-release
-| Type | Description | Stems? |
-|------|-------------|--------|
-| `breakdown_swap` | Outgoing breaks down (LP closes), quiet zone, incoming builds and drops. Maximum narrative tension. | No |
-| `noise_riser_cut` | Synthesized white noise riser (rising HP filter) builds tension over 16 beats → hard cut to incoming. Festival-style. | No |
-
-### Context-Aware Selection
-The transition selector considers:
-- **Set position** — opening (gentle entries) → peak (high-energy techniques) → closing (wind-down)
-- **Energy direction** — rising (buildups, hpf) vs falling (echo out, reverb wash)
-- **Structural features** — if track A ends with a breakdown, use `breakdown_swap`
-- **Harmonic/BPM compatibility** — determines what's physically possible
-- **Variety** — never repeats the same technique consecutively
-
-### DSP Engine
-- **3-band EQ** — Pioneer DJM-style crossover (Low < 250Hz, Mid 250-2500Hz, High > 2500Hz)
-- **Automation curves** — S-curves, exponential ease-in/out, equal-power, stepped, hold-then-drop (not just linear fades)
-- **DJM-style dub echo** — tempo-synced delay with HP on feedback path
-- **Long reverb** — golden-ratio tap spacing with LP damping
-- **Beat stutter** — grain slicing at accelerating subdivisions
-- **Vinyl spinback** — pitch interpolation simulation
-- **Noise riser** — synthesized white noise with rising HP filter
-- **HTDemucs stem separation** — 9.20 dB SDR for surgical bass swaps and vocal isolation
-
-## Benchmarks
-
-| Metric | Target | What It Measures |
-|--------|--------|-----------------|
-| BPM Detection Stability | ≥0.85 | Inter-beat-interval consistency |
-| Key Detection Confidence | ≥0.65 | Krumhansl-Schmuckler correlation |
-| Structural Segmentation | ≥3 segments/track | Novelty-based boundary detection |
-| Cue Point Detection | ≥2 per track | Segment + phrase boundary coverage |
-| Transition Smoothness | ≤2.0 flux ratio | Spectral flux (transition vs steady-state) |
-| Transition Variety | 0 consecutive repeats | Context-aware selection diversity |
-
-## Research Foundation
-
-- **DJtransGAN** (2021): 4-band EQ + fader automation learned from 7,064 real DJ transitions — [arxiv:2110.06525](https://arxiv.org/abs/2110.06525)
-- **Real DJ Statistics** (2020): 94.5% within ±5% BPM, transitions peak at 32-beat phrases — [arxiv:2008.10267](https://arxiv.org/abs/2008.10267)
-- **CUE-DETR** (2024): DJ cue point detection, F1=0.46 at 16-bar phrasing — [arxiv:2407.06823](https://arxiv.org/abs/2407.06823)
-- **All-In-One** (2023): Beat F1=0.958, Segment HR.5F=0.660 — [arxiv:2307.16425](https://arxiv.org/abs/2307.16425)
-- **HTDemucs** (2023): 9.20 dB SDR stem separation — [arxiv:2211.08553](https://arxiv.org/abs/2211.08553)
-- **SongFormer** (2024): Segment boundaries HR.5F=0.703 — [arxiv:2510.02797](https://arxiv.org/abs/2510.02797)
-- **Foote (2000)**: Checkerboard kernel novelty for structural segmentation
-- **Pioneer DJM-900NXS2**: 3-band EQ crossover points (250Hz / 2500Hz), beat FX parameters
+Uses **demucs** stem separation for surgical drum/bass transitions with zero kick clash.