Fix lab plots: EMA time axis, pitch filter, figure sizes (#485)

* Fix lab plots: EMA time axis, pitch filter, figure sizes

- EMA plot: auto-detect (frames, channels) vs (channels, frames)
  orientation — was showing 0.22s instead of 5s
- Pitch contour: filter artifacts above 250Hz (onset detection noise)
- Acoustic analysis plot: figsize increased to (16, 14)
- Comparison plot: figsize increased to (16, 18)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Fix pitch filter: 350Hz threshold, consistent across all plots

- Raised pitch filter from 250Hz to 350Hz (women/children can exceed 250Hz)
- Applied (50, 350) filter consistently in all 3 visualization cells
- Filter SPARC pitch zeros (unvoiced frames) in comparison data

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: satra

tutorials/audio/speech_representations_lab.ipynb

@@ -51,7 +51,43 @@
             "execution_count": null,
             "metadata": {},
             "outputs": [],
-            "source": "import os\nimport urllib.request\nfrom pathlib import Path\n\nimport matplotlib\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport torch\nimport torchaudio.transforms as T\n\nfrom senselab.audio.data_structures import Audio\nfrom senselab.audio.tasks.features_extraction.ppg import (\n    extract_ppg_segments,\n    to_frame_major_posteriorgram,\n    extract_mean_phoneme_durations,\n    extract_ppgs_from_audios,\n    plot_ppg_phoneme_timeline,\n)\nfrom senselab.audio.tasks.features_extraction.sparc import SparcFeatureExtractor\nfrom senselab.audio.tasks.features_extraction.torchaudio import extract_pitch_from_audios\nfrom senselab.audio.tasks.plotting.plotting import play_audio, plot_aligned_panels, plot_specgram, plot_waveform\nfrom senselab.audio.tasks.preprocessing.preprocessing import (\n    downmix_audios_to_mono,\n    resample_audios,\n)\nfrom senselab.utils.data_structures import DeviceType\n\n# Auto-detect device\nif torch.cuda.is_available():\n    device = DeviceType.CUDA\nelif hasattr(torch.backends, \"mps\") and torch.backends.mps.is_available():\n    device = DeviceType.CPU  # MPS not supported by all backends\nelse:\n    device = DeviceType.CPU\nprint(f\"Using device: {device.value}\")"
+            "source": [
+                "import os\n",
+                "import urllib.request\n",
+                "from pathlib import Path\n",
+                "\n",
+                "import matplotlib\n",
+                "import matplotlib.pyplot as plt\n",
+                "import numpy as np\n",
+                "import torch\n",
+                "import torchaudio.transforms as T\n",
+                "\n",
+                "from senselab.audio.data_structures import Audio\n",
+                "from senselab.audio.tasks.features_extraction.ppg import (\n",
+                "    extract_ppg_segments,\n",
+                "    to_frame_major_posteriorgram,\n",
+                "    extract_mean_phoneme_durations,\n",
+                "    extract_ppgs_from_audios,\n",
+                "    plot_ppg_phoneme_timeline,\n",
+                ")\n",
+                "from senselab.audio.tasks.features_extraction.sparc import SparcFeatureExtractor\n",
+                "from senselab.audio.tasks.features_extraction.torchaudio import extract_pitch_from_audios\n",
+                "from senselab.audio.tasks.plotting.plotting import play_audio, plot_aligned_panels, plot_specgram, plot_waveform\n",
+                "from senselab.audio.tasks.preprocessing.preprocessing import (\n",
+                "    downmix_audios_to_mono,\n",
+                "    resample_audios,\n",
+                ")\n",
+                "from senselab.utils.data_structures import DeviceType\n",
+                "\n",
+                "# Auto-detect device\n",
+                "if torch.cuda.is_available():\n",
+                "    device = DeviceType.CUDA\n",
+                "elif hasattr(torch.backends, \"mps\") and torch.backends.mps.is_available():\n",
+                "    device = DeviceType.CPU  # MPS not supported by all backends\n",
+                "else:\n",
+                "    device = DeviceType.CPU\n",
+                "print(f\"Using device: {device.value}\")"
+            ]
         },
         {
             "cell_type": "markdown",
@@ -140,7 +176,43 @@
             "metadata": {},
             "outputs": [],
             "source": [
-                "from pathlib import Path\n\n# Load recorded audio if available, otherwise download sample files\nimport urllib.request\n\nbase_url = \"https://github.com/sensein/senselab/raw/main/src/tests/data_for_testing\"\n\n# Download sample files as fallback\nfor fname in [\"audio_48khz_mono_16bits.wav\", \"audio_48khz_stereo_16bits.wav\"]:\n    dest = Path(\"tutorial_audio_files\") / fname\n    if not dest.exists():\n        urllib.request.urlretrieve(f\"{base_url}/{fname}\", str(dest))\n\n# Sentence 1: use recording if available\nrec1 = Path(\"tutorial_audio_files/recording_1.wav\")\nif rec1.exists() and rec1.stat().st_size > 1000:\n    sentence1 = Audio(filepath=str(rec1.resolve()))\n    print(\"Sentence 1: using your recording\")\nelse:\n    sentence1 = Audio(filepath=os.path.abspath(\"tutorial_audio_files/audio_48khz_mono_16bits.wav\"))\n    print(\"Sentence 1: using sample audio\")\n\n# Sentence 2: use recording if available\nrec2 = Path(\"tutorial_audio_files/recording_2.wav\")\nif rec2.exists() and rec2.stat().st_size > 1000:\n    sentence2_raw = Audio(filepath=str(rec2.resolve()))\n    print(\"Sentence 2: using your recording\")\nelse:\n    sentence2_raw = Audio(filepath=os.path.abspath(\"tutorial_audio_files/audio_48khz_stereo_16bits.wav\"))\n    print(\"Sentence 2: using sample audio\")\n\n# Preprocess: mono + 16kHz\nsentence1_16k = resample_audios(downmix_audios_to_mono([sentence1]), resample_rate=16000)[0]\nsentence2_16k = resample_audios(downmix_audios_to_mono([sentence2_raw]), resample_rate=16000)[0]\n\nprint(f\"Sentence 1: {sentence1_16k.waveform.shape[1] / 16000:.2f}s at 16kHz\")\nprint(f\"Sentence 2: {sentence2_16k.waveform.shape[1] / 16000:.2f}s at 16kHz\")"
+                "from pathlib import Path\n",
+                "\n",
+                "# Load recorded audio if available, otherwise download sample files\n",
+                "import urllib.request\n",
+                "\n",
+                "base_url = \"https://github.com/sensein/senselab/raw/main/src/tests/data_for_testing\"\n",
+                "\n",
+                "# Download sample files as fallback\n",
+                "for fname in [\"audio_48khz_mono_16bits.wav\", \"audio_48khz_stereo_16bits.wav\"]:\n",
+                "    dest = Path(\"tutorial_audio_files\") / fname\n",
+                "    if not dest.exists():\n",
+                "        urllib.request.urlretrieve(f\"{base_url}/{fname}\", str(dest))\n",
+                "\n",
+                "# Sentence 1: use recording if available\n",
+                "rec1 = Path(\"tutorial_audio_files/recording_1.wav\")\n",
+                "if rec1.exists() and rec1.stat().st_size > 1000:\n",
+                "    sentence1 = Audio(filepath=str(rec1.resolve()))\n",
+                "    print(\"Sentence 1: using your recording\")\n",
+                "else:\n",
+                "    sentence1 = Audio(filepath=os.path.abspath(\"tutorial_audio_files/audio_48khz_mono_16bits.wav\"))\n",
+                "    print(\"Sentence 1: using sample audio\")\n",
+                "\n",
+                "# Sentence 2: use recording if available\n",
+                "rec2 = Path(\"tutorial_audio_files/recording_2.wav\")\n",
+                "if rec2.exists() and rec2.stat().st_size > 1000:\n",
+                "    sentence2_raw = Audio(filepath=str(rec2.resolve()))\n",
+                "    print(\"Sentence 2: using your recording\")\n",
+                "else:\n",
+                "    sentence2_raw = Audio(filepath=os.path.abspath(\"tutorial_audio_files/audio_48khz_stereo_16bits.wav\"))\n",
+                "    print(\"Sentence 2: using sample audio\")\n",
+                "\n",
+                "# Preprocess: mono + 16kHz\n",
+                "sentence1_16k = resample_audios(downmix_audios_to_mono([sentence1]), resample_rate=16000)[0]\n",
+                "sentence2_16k = resample_audios(downmix_audios_to_mono([sentence2_raw]), resample_rate=16000)[0]\n",
+                "\n",
+                "print(f\"Sentence 1: {sentence1_16k.waveform.shape[1] / 16000:.2f}s at 16kHz\")\n",
+                "print(f\"Sentence 2: {sentence2_16k.waveform.shape[1] / 16000:.2f}s at 16kHz\")"
             ]
         },
         {
@@ -277,6 +349,9 @@
                 "]\n",
                 "\n",
                 "ema = features1[\"ema\"].numpy()\n",
+                "# EMA shape is (frames, channels) \u2014 detect orientation\n",
+                "if ema.shape[0] > ema.shape[1]:  # (frames, channels)\n",
+                "    ema = ema.T  # transpose to (channels, frames)\n",
                 "n_channels = min(ema.shape[0], len(CHANNEL_NAMES))\n",
                 "n_frames = ema.shape[1]\n",
                 "time_ema = np.arange(n_frames) / 50.0  # SPARC runs at 50 Hz\n",
@@ -444,24 +519,7 @@
             "metadata": {},
             "outputs": [],
             "source": [
-                "# Extract acoustic pitch contour\n",
-                "pitch_result = extract_pitch_from_audios([sentence1_16k], freq_low=80, freq_high=500)\n",
-                "pitch_contour = pitch_result[0][\"pitch\"].squeeze()\n",
-                "\n",
-                "# Build time axis (default hop length for torchaudio pitch detection)\n",
-                "hop_length = 160  # default for 16 kHz\n",
-                "time_pitch = torch.arange(len(pitch_contour)) * hop_length / 16000\n",
-                "\n",
-                "# Plot only voiced frames (pitch > 0)\n",
-                "plt.figure(figsize=(12, 4))\n",
-                "voiced = pitch_contour > 0\n",
-                "plt.scatter(time_pitch[voiced].numpy(), pitch_contour[voiced].numpy(), s=2, c=\"blue\")\n",
-                "plt.xlabel(\"Time (seconds)\")\n",
-                "plt.ylabel(\"Frequency (Hz)\")\n",
-                "plt.title(\"Acoustic Pitch (F0) Contour\")\n",
-                "plt.grid(True, alpha=0.3)\n",
-                "plt.tight_layout()\n",
-                "plt.show()"
+                "# Extract acoustic pitch contour\npitch_result = extract_pitch_from_audios([sentence1_16k], freq_low=80, freq_high=500)\npitch_contour = pitch_result[0][\"pitch\"].squeeze()\n\n# Build time axis (default hop length for torchaudio pitch detection)\nhop_length = 160  # default for 16 kHz\ntime_pitch = torch.arange(len(pitch_contour)) * hop_length / 16000\n\n# Plot only voiced frames (pitch > 0)\nplt.figure(figsize=(12, 4))\nvoiced = (pitch_contour > 50) & (pitch_contour < 350)  # filter artifacts\nplt.scatter(time_pitch[voiced].numpy(), pitch_contour[voiced].numpy(), s=2, c=\"blue\")\nplt.xlabel(\"Time (seconds)\")\nplt.ylabel(\"Frequency (Hz)\")\nplt.title(\"Acoustic Pitch (F0) Contour\")\nplt.grid(True, alpha=0.3)\nplt.tight_layout()\nplt.show()"
             ]
         },
         {
@@ -535,7 +593,9 @@
             "execution_count": null,
             "metadata": {},
             "outputs": [],
-            "source": "# Standard spectrogram params: 10ms window, 5ms hop at 16kHz\nspec_params = {\"n_fft\": 256, \"hop_length\": 80, \"win_length\": 160}\n\n# Get PPG segments for plotting\nframe_major = to_frame_major_posteriorgram(ppg1)\nsegments = extract_ppg_segments(sentence1_16k, frame_major)\n\n# Convert PPG segments to plot_aligned_panels format\nppg_segments = [\n    {\"label\": seg[\"phoneme\"], \"start\": seg[\"start_seconds\"], \"end\": seg[\"start_seconds\"] + seg[\"duration_seconds\"]}\n    for seg in segments\n]\n\n# Build pitch overlay data\nvoiced_mask = pitch_contour > 0\npitch_data = [(time_pitch[voiced_mask], pitch_contour[voiced_mask], \"F0\", \"steelblue\")]\n\nplot_aligned_panels(\n    sentence1_16k,\n    [\n        {\"type\": \"waveform\"},\n        {\"type\": \"spectrogram\", \"mel\": True},\n        {\"type\": \"features\", \"data\": pitch_data},\n        {\"type\": \"segments\", \"segments\": ppg_segments},\n    ],\n    title=\"Acoustic Analysis: Waveform + Spectrogram + Pitch + PPG\",\n    spectrogram_params=spec_params,\n)\nplt.show()"
+            "source": [
+                "# Standard spectrogram params: 10ms window, 5ms hop at 16kHz\nspec_params = {\"n_fft\": 256, \"hop_length\": 80, \"win_length\": 160}\n\n# Get PPG segments for plotting\nframe_major = to_frame_major_posteriorgram(ppg1)\nsegments = extract_ppg_segments(sentence1_16k, frame_major)\n\n# Convert PPG segments to plot_aligned_panels format\nppg_segments = [\n    {\"label\": seg[\"phoneme\"], \"start\": seg[\"start_seconds\"], \"end\": seg[\"start_seconds\"] + seg[\"duration_seconds\"]}\n    for seg in segments\n]\n\n# Build pitch overlay data\nvoiced_mask = (pitch_contour > 50) & (pitch_contour < 350)\npitch_data = [(time_pitch[voiced_mask], pitch_contour[voiced_mask], \"F0\", \"steelblue\")]\n\nplot_aligned_panels(\n    sentence1_16k,\n    [\n        {\"type\": \"waveform\"},\n        {\"type\": \"spectrogram\", \"mel\": True},\n        {\"type\": \"features\", \"data\": pitch_data},\n        {\"type\": \"segments\", \"segments\": ppg_segments},\n    ],\n    title=\"Acoustic Analysis: Waveform + Spectrogram + Pitch + PPG\",\n    spectrogram_params=spec_params,\n    figsize=(16, 14),\n)\nplt.show()"
+            ]
         },
         {
             "cell_type": "markdown",
@@ -564,7 +624,9 @@
             "execution_count": null,
             "metadata": {},
             "outputs": [],
-            "source": "# Get PPG segments for the comparison\nframe_major = to_frame_major_posteriorgram(ppg1)\nsegments = extract_ppg_segments(sentence1_16k, frame_major)\nppg_segments = [\n    {\"label\": seg[\"phoneme\"], \"start\": seg[\"start_seconds\"], \"end\": seg[\"start_seconds\"] + seg[\"duration_seconds\"]}\n    for seg in segments\n]\n\n# Build SPARC pitch and acoustic pitch as feature overlays\nsparc_pitch_time = np.arange(len(pitch_sparc)) / 50.0\nvoiced_mask = pitch_contour > 0\n\npitch_comparison_data = [\n    (sparc_pitch_time, pitch_sparc, \"SPARC pitch\", \"coral\"),\n    (time_pitch[voiced_mask], pitch_contour[voiced_mask], \"Acoustic F0\", \"steelblue\"),\n]\n\nloudness_data = [(loud_time, loudness_sparc, \"SPARC Loudness\", \"coral\")]\n\n# Selected EMA trajectories\nselected_channels = [(\"TT_y\", 1), (\"TB_y\", 3), (\"UL_y\", 7), (\"LL_y\", 9)]\nema_data = []\nfor name, ch_idx in selected_channels:\n    artic = name.split(\"_\")[0]\n    color_map = {\n        \"TT\": \"#43B96B\",\n        \"TB\": \"#3EADD8\",\n        \"UL\": \"#EE3A5B\",\n        \"LL\": \"#FFD155\",\n    }\n    signal = ema[ch_idx] - ema[ch_idx].mean()\n    ema_data.append((time_ema, signal, name, color_map.get(artic, \"gray\")))\n\nplot_aligned_panels(\n    sentence1_16k,\n    [\n        {\"type\": \"waveform\"},\n        {\"type\": \"features\", \"data\": pitch_comparison_data},\n        {\"type\": \"features\", \"data\": loudness_data},\n        {\"type\": \"features\", \"data\": ema_data},\n        {\"type\": \"segments\", \"segments\": ppg_segments},\n    ],\n    title=\"Acoustic vs Articulatory Representations (side by side)\",\n    spectrogram_params=spec_params,\n)\nplt.show()"
+            "source": [
+                "# Get PPG segments for the comparison\nframe_major = to_frame_major_posteriorgram(ppg1)\nsegments = extract_ppg_segments(sentence1_16k, frame_major)\nppg_segments = [\n    {\"label\": seg[\"phoneme\"], \"start\": seg[\"start_seconds\"], \"end\": seg[\"start_seconds\"] + seg[\"duration_seconds\"]}\n    for seg in segments\n]\n\n# Build SPARC pitch and acoustic pitch as feature overlays\nsparc_pitch_time = np.arange(len(pitch_sparc)) / 50.0\nvoiced_mask = (pitch_contour > 50) & (pitch_contour < 350)\n\npitch_comparison_data = [\n    (sparc_pitch_time, pitch_sparc, \"SPARC pitch\", \"coral\"),\n    (time_pitch[voiced_mask], pitch_contour[voiced_mask], \"Acoustic F0\", \"steelblue\"),\n]\n\nloudness_data = [(loud_time, loudness_sparc, \"SPARC Loudness\", \"coral\")]\n\n# Selected EMA trajectories\nselected_channels = [(\"TT_y\", 1), (\"TB_y\", 3), (\"UL_y\", 7), (\"LL_y\", 9)]\nema_data = []\nfor name, ch_idx in selected_channels:\n    artic = name.split(\"_\")[0]\n    color_map = {\n        \"TT\": \"#43B96B\",\n        \"TB\": \"#3EADD8\",\n        \"UL\": \"#EE3A5B\",\n        \"LL\": \"#FFD155\",\n    }\n    signal = ema[ch_idx] - ema[ch_idx].mean()\n    ema_data.append((time_ema, signal, name, color_map.get(artic, \"gray\")))\n\nplot_aligned_panels(\n    sentence1_16k,\n    [\n        {\"type\": \"waveform\"},\n        {\"type\": \"features\", \"data\": pitch_comparison_data},\n        {\"type\": \"features\", \"data\": loudness_data},\n        {\"type\": \"features\", \"data\": ema_data},\n        {\"type\": \"segments\", \"segments\": ppg_segments},\n    ],\n    title=\"Acoustic vs Articulatory Representations (side by side)\",\n    spectrogram_params=spec_params,\n    figsize=(16, 18),\n)\nplt.show()"
+            ]
         },
         {
             "cell_type": "markdown",