Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+example_melody.wav filter=lfs diff=lfs merge=lfs -text
+example_piano.wav filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,31 @@
 ---
-title: Muscriptor Music Transcription
-emoji: 🔥
-colorFrom: gray
-colorTo: green
+title: MuScriptor Music Transcription
+emoji: 🎵
+colorFrom: red
+colorTo: blue
 sdk: gradio
 sdk_version: 6.20.0
-python_version: '3.12'
 app_file: app.py
-pinned: false
+short_description: Convert music audio to MIDI with MuScriptor
+python_version: "3.12"
+startup_duration_timeout: 30m
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# MuScriptor Music Transcription
+
+This Space demonstrates [MuScriptor](https://huggingface.co/MuScriptor/muscriptor-medium), a ~300M parameter multi-instrument automatic music transcription (AMT) model developed by Mirelo x Kyutai.
+
+Upload a music recording (any genre, multiple instruments) and the model will transcribe it into a downloadable MIDI file with per-note onset, offset, pitch, and instrument information.
+
+## How it works
+
+1. Upload an audio file (WAV, MP3, FLAC, etc.) or record from microphone
+2. Optionally select which instruments are present to improve accuracy
+3. Click "Transcribe" to get a downloadable MIDI file
+
+## Model details
+
+- **Architecture**: Decoder-only Transformer (dim=1024, 16 heads, 24 layers)
+- **Input**: 16 kHz mono audio → mel-spectrogram (512 mel bins)
+- **Output**: MIDI-like token sequence using the MT3_FULL_PLUS taxonomy (36 instrument subgroups)
+- **License**: CC-BY-NC 4.0 (weights) / MIT (code)

app.py

@@ -0,0 +1,164 @@
+import spaces  # MUST come before any torch / CUDA-touching import
+import torch
+import gradio as gr
+import tempfile
+import os
+import time
+import sys
+import io
+
+from muscriptor.transcription_model import TranscriptionModel
+from muscriptor.events import NoteStartEvent, NoteEndEvent, ProgressEvent
+from muscriptor.tokenizer.mt3 import MT3_FULL_PLUS_GROUP_NAMES
+
+# Load the model at module scope — ZeroGPU intercepts .to("cuda") and packs
+# weights to disk, streaming them into VRAM on the first @spaces.GPU call.
+print("[muscriptor-space] Loading model...", file=sys.stderr, flush=True)
+t0 = time.perf_counter()
+model = TranscriptionModel.load_model("medium", device="cuda")
+print(f"[muscriptor-space] Model loaded in {time.perf_counter() - t0:.1f}s", file=sys.stderr, flush=True)
+
+# Build the instrument choices list (sorted by group ID for stable ordering)
+INSTRUMENT_CHOICES = sorted(MT3_FULL_PLUS_GROUP_NAMES.keys(), key=lambda k: MT3_FULL_PLUS_GROUP_NAMES[k])
+
+CSS = """
+#col-container { max-width: 1100px; margin: 0 auto; }
+.dark .gradio-container { color: var(--body-text-color); }
+"""
+
+
+@spaces.GPU(duration=120)
+def transcribe_audio(
+    audio_path: str,
+    instruments: list[str] | None,
+    use_sampling: bool,
+    temperature: float,
+    progress=gr.Progress(track_tqdm=True),
+):
+    """Transcribe an audio recording into a downloadable MIDI file.
+
+    Upload any music recording (multi-instrument, any genre) and MuScriptor
+    will convert it into a MIDI file with per-note onset, offset, pitch, and
+    instrument information.
+
+    Args:
+        audio_path: Path to the uploaded audio file (wav, mp3, flac, etc.).
+        instruments: Optional list of instrument group names to condition the
+            transcription (improves coherence when you know which instruments
+            are present). Leave empty for automatic (unconditioned) transcription.
+        use_sampling: If True, use stochastic sampling instead of greedy decoding.
+        temperature: Sampling temperature (only used when use_sampling is True).
+    """
+    if audio_path is None:
+        return None, "Please upload an audio file first."
+
+    t0 = time.perf_counter()
+
+    # Run the transcription — returns MIDI bytes
+    try:
+        midi_bytes = model.transcribe_to_midi(
+            audio_path,
+            instruments=instruments if instruments else None,
+            use_sampling=use_sampling,
+            temperature=temperature,
+        )
+    except Exception as e:
+        return None, f"Transcription failed: {e}"
+
+    elapsed = time.perf_counter() - t0
+
+    # Write to a temporary file for download
+    tmp = tempfile.NamedTemporaryFile(suffix=".mid", delete=False)
+    tmp.write(midi_bytes)
+    tmp.close()
+
+    # Also collect note statistics for the summary
+    note_count = 0
+    instrument_set = set()
+    try:
+        from mido import MidiFile
+        midi = MidiFile(tmp.name)
+        for track in midi.tracks:
+            for msg in track:
+                if msg.type == "note_on" and msg.velocity > 0:
+                    note_count += 1
+                    instrument_set.add(msg.program if not msg.is_meta else 0)
+    except Exception:
+        pass
+
+    summary = (
+        f"Transcription complete in {elapsed:.1f}s. "
+        f"Found {note_count} notes across {len(instrument_set)} instrument program(s). "
+        f"Download the MIDI file below."
+    )
+
+    return tmp.name, summary
+
+
+with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.Markdown(
+            """
+            # MuScriptor — Music Transcription (Audio → MIDI)
+
+            Upload a music recording and get a downloadable MIDI file with transcribed notes.
+            MuScriptor is a ~300M parameter multi-instrument automatic music transcription model
+            developed by [Mirelo](https://www.mirelo.ai/) x [Kyutai](https://kyutai.org/).
+
+            [Model card](https://huggingface.co/MuScriptor/muscriptor-medium) · [Code](https://github.com/muscriptor/muscriptor) · [Audio samples](https://muscriptor.github.io)
+            """
+        )
+
+        with gr.Row():
+            audio_input = gr.Audio(
+                label="Upload or record audio",
+                type="filepath",
+                sources=["upload", "microphone"],
+            )
+
+        with gr.Accordion("Advanced settings", open=False):
+            instrument_checkbox = gr.CheckboxGroup(
+                choices=INSTRUMENT_CHOICES,
+                value=[],
+                label="Instrument conditioning (optional)",
+                info="Select instruments present in the audio to improve transcription accuracy. Leave empty for automatic detection.",
+            )
+            use_sampling = gr.Checkbox(
+                label="Use sampling (stochastic decoding)",
+                value=False,
+                info="If enabled, uses temperature-based sampling instead of greedy decoding.",
+            )
+            temperature = gr.Slider(
+                label="Temperature",
+                minimum=0.1,
+                maximum=2.0,
+                value=1.0,
+                step=0.1,
+                info="Sampling temperature (only used when sampling is enabled).",
+            )
+
+        transcribe_btn = gr.Button("Transcribe", variant="primary")
+
+        midi_output = gr.File(label="Download MIDI file")
+        summary_output = gr.Textbox(label="Summary", interactive=False)
+
+    transcribe_btn.click(
+        fn=transcribe_audio,
+        inputs=[audio_input, instrument_checkbox, use_sampling, temperature],
+        outputs=[midi_output, summary_output],
+        api_name="transcribe",
+    )
+
+    gr.Examples(
+        examples=[
+            ["example_piano.wav", [], False, 1.0],
+            ["example_melody.wav", [], False, 1.0],
+        ],
+        inputs=[audio_input, instrument_checkbox, use_sampling, temperature],
+        outputs=[midi_output, summary_output],
+        fn=transcribe_audio,
+        cache_examples=True,
+        cache_mode="lazy",
+    )
+
+demo.launch(mcp_server=True)

example_melody.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac1baf0ca3cd75e9e91fff3fa03980eb2696def24ac36dd78786beffc9d87960
+size 256044

example_piano.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c69639c8efc102188d0f55cbc7b05528339e5a4b38062dccf7a99bb71d5efac
+size 160044

requirements.txt

@@ -0,0 +1,6 @@
+git+https://github.com/muscriptor/muscriptor.git
+numpy>=1.24
+einops>=0.4
+mido>=1.3
+safetensors>=0.4
+soundfile>=0.14.0