pulsevad 0.1.0__tar.gz

pulsevad-0.1.0/.claude-flow/policy/state.json

@@ -0,0 +1,52 @@
+{
+  "version": 1,
+  "mode": "legacy",
+  "rules": [],
+  "budgets": [],
+  "usage": [],
+  "approvals": [],
+  "receipts": [
+    {
+      "payload": {
+        "receiptId": "sha256:eb3c2b2101ffe1b64c3bda8261a941075647bf8718d86d9ee27e86c59b7e2b91",
+        "previousReceiptHash": null,
+        "sequence": 0,
+        "issuedAt": 1788261872543,
+        "request": {
+          "identity": {
+            "id": "legacy-cli",
+            "type": "legacy"
+          },
+          "action": {
+            "type": "mcp.tool.call",
+            "resource": "hooks_session-end",
+            "tool": "hooks_session-end",
+            "server": "ruflo",
+            "network": false,
+            "destructive": false
+          },
+          "context": {
+            "metadata": {
+              "inputDigest": "sha256:9ddccf94f2e2940231b45b399e38a0221ef77ec3a79b1a042acf0e2e4aa01c36"
+            },
+            "now": 1788261872543
+          },
+          "requestId": "1a79a4fa-6deb-4cc4-b3fc-e7b9dffe9a52"
+        },
+        "decision": {
+          "requestId": "1a79a4fa-6deb-4cc4-b3fc-e7b9dffe9a52",
+          "outcome": "allowed",
+          "enforcedOutcome": "allowed",
+          "mode": "legacy",
+          "reason": "legacy-default-allow",
+          "matchedRules": [],
+          "obligations": []
+        },
+        "policyHash": "sha256:00b63b2f746eb8b9095bbd5e1dbae4ddd74fd92a6740b72f59f4069fbfda4737"
+      },
+      "hash": "sha256:0c1ec83309dd286666d559ab58a7e3a0ef8cc07aa720bf2563a0099cacf3d9f9"
+    }
+  ],
+  "migratedAt": 1788261872543,
+  "migratedFrom": "pre-ADR-324; capabilities=none-detected"
+}

pulsevad-0.1.0/.github/workflows/publish.yml

@@ -0,0 +1,26 @@
+name: Publish to PyPI
+
+on:
+  release:
+    types: [published]
+  workflow_dispatch:
+
+jobs:
+  pypi-publish:
+    name: Upload release to PyPI
+    runs-on: ubuntu-latest
+    permissions:
+      id-token: write  # MANDATORY for PyPI Trusted Publishing
+      contents: read
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@v4
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
+
+      - name: Build package distributions
+        run: uv build
+
+      - name: Publish to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1

pulsevad-0.1.0/.gitignore

@@ -0,0 +1,24 @@
+__pycache__/
+*.pyc
+data/raw/*
+data/labels/*
+data/cache/*
+*.pth
+*.onnx
+*.tflite
+.venv/
+dist/
+build/
+.pytest_cache/
+!data/raw/.gitkeep
+!data/labels/.gitkeep
+!data/cache/.gitkeep
+.agents/
+data/runs/*
+!data/runs/.gitkeep
+specs/
+*.pdf
+*build_plan*
+!pulsevad/data/*
+!pulsevad/data/**
+

pulsevad-0.1.0/.python-version

@@ -0,0 +1 @@
+3.11

pulsevad-0.1.0/ATTRIBUTION.md

@@ -0,0 +1,30 @@
+# Attribution
+
+PulseVAD is trained exclusively on commercially clean, permissively licensed datasets and
+tools. This document acknowledges the sources used.
+
+## Datasets
+
+- **LibriSpeech train-clean-100** — CC BY 4.0 — Vassil Panayotov, Guoguo Chen, Daniel Povey,
+  Sanjeev Khudanpur. Primary speech corpus.
+- **Common Voice** — CC0 (Public Domain) — Mozilla. Multilingual speech expansion.
+- **Multilingual LibriSpeech (MLS)** — CC BY 4.0. Multilingual speech expansion.
+- **VoxLingua107** — CC BY 4.0. Multilingual speech expansion.
+- **MUSAN (Noise / Music)** — CC BY 4.0 — David Snyder, Guoguo Chen, Daniel Povey. Noise
+  augmentation and background noise.
+- **DNS Challenge Noise (Interspeech 2020)** — CC BY 4.0 / CC0 subset — Chandan K. A. Reddy et
+  al. Acoustic interference augmentation.
+- **Synthetic Wind Noise (Mirabilii 2022)** — Open source algorithm, synthesized locally.
+  Airflow simulation at -5 dB SNR.
+
+## Tools
+
+- **Silero-VAD (v5/v6)** — MIT License — Silero Team. Used only as an automated self-labeling
+  engine; no Silero-released labeled dataset or weights are used or shipped.
+
+## Explicitly Excluded (Non-Commercial, Do NOT Use)
+
+- Silero Released Labeled Dataset — CC BY-NC-SA 4.0 — never downloaded or used.
+- LibriVAD Dataset — CC BY-NC-SA 4.0 — never downloaded or used.
+- Official kiloVAD Checkpoints — CC BY-NC 4.0 — referenced only as a numerical unit-test
+  reference, never shipped or trained upon.

pulsevad-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Aydin Adnan
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

pulsevad-0.1.0/PKG-INFO

@@ -0,0 +1,19 @@
+Metadata-Version: 2.5
+Name: pulsevad
+Version: 0.1.0
+Summary: Ultra-tiny streaming VAD (kiloVAD clean-room reimplementation)
+License-File: LICENSE
+Requires-Python: >=3.11
+Requires-Dist: matplotlib>=3.8.0
+Requires-Dist: modal>=1.5.5
+Requires-Dist: numpy<2.0.0,>=1.24.0
+Requires-Dist: onnx>=1.15.0
+Requires-Dist: onnxruntime>=1.16.0
+Requires-Dist: pyroomacoustics>=0.7.5
+Requires-Dist: scikit-learn>=1.3.0
+Requires-Dist: scipy>=1.11.0
+Requires-Dist: soundfile>=0.12.1
+Requires-Dist: torch-pruning>=1.4.0
+Requires-Dist: torch>=2.1.0
+Requires-Dist: torchaudio>=2.1.0
+Requires-Dist: tqdm>=4.66.0

pulsevad-0.1.0/README.md

@@ -0,0 +1,170 @@
+# pulsevad
+
+2,118 parameters. 2.1 KB as INT8. strictly causal. zero future context. runs on microcontrollers that choke on silero.
+
+an ultra-compact, commercially clean voice activity detector built from scratch, inspired by the kiloVAD architecture in [arXiv:2607.25870v1](https://arxiv.org/abs/2607.25870) (*"VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment"*, INTERSPEECH 2026).
+
+![benchmarks](docs/assets/comparison_graph.png)
+
+---
+
+## installation
+
+### pip / uv
+```bash
+# coming soon
+```
+
+to run or reproduce the training, pruning, and cloud evaluation pipeline on Modal yourself, see [REPRODUCE.md](REPRODUCE.md).
+
+---
+
+## what this actually is
+
+if you've ever tried running a modern deep learning voice activity detector on a real embedded target (think an ARM Cortex-M0+ or M4 with 32 KB of RAM), you know the options suck. silero is fantastic for servers and desktop apps, but it weighs 545,000 parameters (~2.2 MB) and demands millions of MACs per inference. other tiny models in academic papers either rely on non-causal lookahead (cheating by looking 600 ms into the future), require exotic activation functions that don't exist in CMSIS-NN, or use non-commercial research licenses.
+
+pulsevad takes raw 16 kHz mono audio, computes a 64-channel log-mel spectrogram over a 200 ms causal window, and runs a depthwise-separable 1D CNN pruned down to **2,118 parameters**. 
+
+quantized with round-to-nearest INT8, the entire weight payload is **2.1 KB**. it ships as a single drop-in C header (`pulsevad_weights.h`) and standard ONNX graphs.
+
+---
+
+## why building this was hell (and how it actually works)
+
+you cannot train a 2,118 parameter network from scratch on noisy audio. it gets stuck in terrible local minima and predicts pure noise 100% of the time.
+
+here is how we got it to work:
+
+1. **the 81k teacher**: we first trained an 81,090-parameter CNN backbone on LibriSpeech augmented with synthetic room impulse responses (pyroomacoustics), synthetic wind profiles, and heavy background noise from MUSAN at -10 dB to +10 dB SNR.
+2. **commercially clean self-labeling**: instead of using non-commercial academic labels (like LibriVAD or AVA CC-BY-NC splits), we labeled 28,539 LibriSpeech files using Silero-VAD under an MIT license, using a 0.50/0.35 hysteresis state machine quantized to a strict 10 ms grid. 100% permissive commercial data only.
+3. **DepGraph structured pruning**: uniform pruning collapses at sub-3k params. we used dependency-graph magnitude pruning to identify coupled channel dependencies across depthwise and pointwise layers, carving out the exact 2.1k channel spec.
+4. **knowledge distillation**: we fine-tuned the 2.1k student under the frozen 81k teacher using KL divergence with temperature scaling and cosine learning rate decay.
+5. **the silent bias trap**: LibriSpeech is ~78% active speech. a distilled model naturally inherits a positive prior (+1.4 logit shift), which causes it to trigger on pure air conditioning noise (yielding a disastrous 25% false positive rate on pure noise). we implemented quantile-based prior bias calibration to shift the linear classifier bias, crushing pure-noise FPR to **3.5%** while keeping ROC-AUC strictly invariant.
+6. **post-training quantization (PTQ)**: batchnorm layers were mathematically folded into conv weights before calibration. using symmetric per-channel weight scaling and per-tensor activation scaling, the AUC gap between FP32 and INT8 is under **0.001**.
+
+---
+
+## cross-vad benchmark: size, compute & latency
+
+we benchmarked pulsevad against standard embedded baselines and measured silero-vad v5 on identical causal 200 ms audio chunks:
+
+| model | params | footprint | MACs / 200 ms | input latency | commercial license |
+| :--- | :---: | :---: | :---: | :---: | :---: |
+| **PulseVAD (81k teacher)** *(measured)* | 81,090 | 324 KB FP32 / 81 KB INT8 | 1.66M | 200 ms | **YES (MIT)** |
+| **PulseVAD (2.1k ship)** *(measured)* | **2,118** | **8.5 KB FP32 / 2.1 KB INT8** | **44,000** | **200 ms** | **YES (MIT)** |
+| **Silero-VAD (v6)** *(measured)* | 309,000 | ~1.2 MB | >8M | 32 ms | YES (MIT) |
+| **Silero-VAD (v5)** *(measured)* | 545,000 | ~2.2 MB | >10M | 32 ms | YES (MIT) |
+| **MarbleNet** *(measured)* | 91,000 | ~370 KB | >2.0M | 630 ms | non-commercial (NS) |
+| **AtomicVAD** *[cited]* | 300 | ~1.2 KB | 6,000 | 630 ms | non-commercial (custom GGCU) |
+| **TinyVAD** *[cited]* | 11,600 | n/a | ~80,000 | 630 ms | non-causal (87.5% lookahead) |
+| **ResectNet** *[cited]* | 4,500 | n/a | n/a | 200 ms | non-commercial |
+
+---
+
+## acoustic evaluation: measured AUC on held-out test sets
+
+evaluated causally with 0% overlap across 2,000 audio windows per category on the exact same audio:
+
+| category | PulseVAD 2.1k INT8 | PulseVAD 2.1k FP32 | PulseVAD 81k Teacher | Silero-VAD v6 (measured) | Silero-VAD v5 (measured) | MarbleNet (measured) |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: |
+| **clean speech** | 0.976 | 0.977 | **0.989** | 0.988 | 0.990 | 0.970 |
+| **windy / reverb** | **0.938** | 0.937 | **0.985** | 0.934 | 0.960 | 0.910 |
+| **DNS synthetic noise** | **0.891** | 0.891 | **0.943** | 0.873 | 0.913 | 0.903 |
+| **speech + noise (0-20 dB)** | 0.903 | 0.904 | **0.966** | 0.916 | 0.944 | 0.924 |
+| **pure noise (FPR@95, gate <5%)**| **0.035 (3.5%)** | 0.040 (4.0%) | 0.040 (4.0%) | 0.004 (0.4%) | 0.017 (1.7%) | 0.024 (2.4%) |
+
+*silero v5 and v6 were streamed in 512-sample causal chunks with hidden states reset per window. marblenet was evaluated using its 80-channel mel frontend in 11-frame output aggregations.*
+
+---
+
+## multilingual & indian language benchmark
+
+we tested generalization on in-the-wild speech across 10 diverse languages from Google FLEURS mixed with realistic MUSAN noise at 0 to 20 dB SNR. this includes 4 major Indian languages (Hindi, Tamil, Telugu, Bengali):
+
+![multilingual](docs/assets/multilingual_graph.png)
+
+| language | PulseVAD 2.1k INT8 | PulseVAD 2.1k FP32 | PulseVAD 81k Teacher | Silero-VAD v5 (545k) |
+| :--- | :---: | :---: | :---: | :---: |
+| **English (US)** | **0.882** | 0.882 | 0.899 | 0.725 |
+| **Spanish (LatAm)** | **0.866** | 0.866 | 0.884 | 0.846 |
+| **French** | **0.819** | 0.820 | 0.876 | 0.812 |
+| **German** | **0.854** | 0.855 | 0.872 | 0.819 |
+| **Mandarin Chinese** | **0.826** | 0.826 | 0.864 | 0.777 |
+| **Japanese** | 0.779 | 0.777 | **0.815** | 0.794 |
+| **Hindi (India)** | **0.872** | 0.872 | 0.923 | 0.857 |
+| **Tamil (India)** | **0.881** | 0.882 | 0.926 | 0.879 |
+| **Telugu (India)** | **0.903** | 0.904 | 0.924 | 0.873 |
+| **Bengali (India)** | 0.843 | 0.843 | **0.884** | 0.845 |
+| **Macro Average** | **0.852** | **0.853** | **0.887** | **0.823** |
+
+in noisy speech conditions, pulsevad's mel frontend and dilated convolutions beat silero v5 across 8 out of 10 languages, outperforming it on average by **+0.029 AUC** while using **257x fewer parameters**.
+
+---
+
+## real talk: advantages vs disadvantages
+
+no model is magic. here is the honest breakdown of when you should use pulsevad and when you should not.
+
+### advantages
+- **runs anywhere**: 2.1 KB fits in L1 cache or tiny MCU SRAM without external DRAM.
+- **zero dependencies**: pure C array weights (`pulsevad_weights.h`) or standard ONNX. no PyTorch, no heavy runtime, no recurrent state tensors to track.
+- **tough on noise**: holds up remarkably well against wind, reverberation, and babble noise because it was trained with aggressive augmentations.
+- **100% commercially permissive**: clean MIT license with no non-commercial viral traps.
+- **strictly causal**: 0 ms lookahead. what happens in the future stays in the future.
+
+### disadvantages
+- **200 ms window granularity**: silero streams in 32 ms sub-chunks. if you need instantaneous 30 ms word-boundary cuts for live transcription, pulsevad's 200 ms input buffer has higher initial buffering latency.
+- **clean speech ceiling**: on pristine studio speech with zero background noise, silero's hundreds of thousands of parameters give it a higher ceiling (0.988–0.990 vs 0.976 AUC).
+- **requires mel frontend**: pulsevad expects 64 log-mel bins. you need an FFT + mel filterbank implementation on your target device (though standard CMSIS-DSP covers this easily).
+
+### head-to-head: vs marblenet & silero (v5 vs v6)
+- **vs MarbleNet (91k params)**: MarbleNet was NVIDIA's lightweight VAD for NeMo. at 91,000 parameters and >2M MACs, it incurs 630 ms input latency and carries a non-commercial license. PulseVAD 2.1k is **43x smaller**, **3.1x lower latency**, and beats MarbleNet on clean speech (**0.976 vs 0.970 AUC**) and windy audio (**0.938 vs 0.910 AUC**).
+- **vs Silero v5 (545k) & v6 (309k)**: Silero v6 trimmed parameters from 545k to 309k (~1.2 MB). while both Silero versions excel on clean studio audio (0.988–0.990 AUC), PulseVAD 2.1k INT8 beats Silero v6 on windy audio (**0.938 vs 0.934**) and DNS synthetic noise (**0.891 vs 0.873**) while being **145x smaller** than v6 and **257x smaller** than v5.
+
+---
+
+## under the hood
+
+```
+16 kHz mono audio (3,200 samples = 200 ms)
+      │
+      ▼
+[ pre-emphasis (0.97) + z-norm ]
+      │
+      ▼
+[ 64-bin log-mel filterbank (21 frames x 64 bins) ]
+      │
+      ▼
+[ conv1d adapter: 64 -> 16 ]
+      │
+      ▼
+[ depthwise-separable block 1: k=11, ch=16 ]
+      │
+      ▼
+[ depthwise-separable block 2: k=17, ch=16 ]
+      │
+      ▼
+[ dilated depthwise block 3: k=29, dilation=2, ch=16 ]
+      │
+      ▼
+[ global average pooling -> linear (16 -> 2) ]
+      │
+      ▼
+[ calibrated prior bias shift (-1.40) ] -> P(speech)
+```
+
+---
+
+## references & attribution
+
+- **kiloVAD Paper**: Stephen Bauer, Sheila Seidel, Shanza Iftikhar, Scott Veidenheimer, Gorkem Ulkar. *"VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment"*, [arXiv:2607.25870v1](https://arxiv.org/abs/2607.25870), INTERSPEECH 2026. (inspiration for ultra-tiny CNN VAD and structural pruning targets).
+- **Silero-VAD**: [snakers4/silero-vad](https://github.com/snakers4/silero-vad) (MIT License) used as the teacher state-machine labeling tool and evaluation baseline.
+- **Google FLEURS**: [google/fleurs](https://huggingface.co/datasets/google/fleurs) for multilingual speech evaluation.
+- **LibriSpeech & OpenSLR**: Vassil Panayotov et al., [OpenSLR 12](https://www.openslr.org/12/) (CC BY 4.0).
+- **MUSAN Corpus**: David Snyder et al., [OpenSLR 17](https://www.openslr.org/17/) (CC BY 4.0).
+
+---
+
+## license
+
+MIT License. see [LICENSE](LICENSE) for details.

pulsevad-0.1.0/REPRODUCE.md

@@ -0,0 +1,110 @@
+# reproduce pulsevad end-to-end
+
+this guide documents how to reproduce the entire pulsevad training, pruning, quantization, export, and cross-vad benchmarking pipeline on modal.
+
+---
+
+## prerequisites
+
+1. install `uv` (fast python package manager):
+   ```bash
+   curl -LsSf https://astral.sh/uv/install.sh | sh
+   ```
+2. clone this repo and sync dependencies:
+   ```bash
+   git clone https://github.com/AydinAdnan/PulseVAD.git
+   cd PulseVAD
+   uv sync
+   ```
+3. authenticate with modal:
+   ```bash
+   uv run modal setup
+   ```
+
+---
+
+## pipeline execution steps
+
+### 1. download training corpora
+pulls librispeech dev-clean (speech) and musan (noise, music, speech) onto the `pulsevad-data` cloud volume:
+```bash
+uv run modal run modal_app.py::download
+```
+
+### 2. generate teacher speech labels
+labels audio with silero vad hysteresis state machine into canonical 10 ms grid manifests:
+```bash
+uv run modal run modal_app.py::label
+```
+
+### 3. synthesize augmented window cache
+generates 3.4m augmented 200 ms windows with noise mixing, simulated rirs, and noise-only rejection samples:
+```bash
+uv run modal run modal_app.py::build_cache --seed 0
+```
+
+### 4. train unpruned teacher model (81k params)
+trains the unpruned 81k cnn teacher on the augmented cache for 40 epochs with cosine annealing:
+```bash
+uv run modal run modal_app.py::train --seed 0
+```
+
+### 5. structured pruning & self-distillation (2.1k student)
+applies depgraph magnitude pruning to collapse channels to the exact 2,118 parameter spec and runs 8-epoch self-distillation with kl divergence:
+```bash
+uv run modal run modal_app.py::distill --seed 0
+```
+
+### 6. batchnorm folding, int8 quantization & onnx / c export
+folds batchnorm into conv weights, runs rtn int8 quantization with activation calibration, verifies the <= 0.002 auc gap, and exports onnx pairs and `pulsevad_weights.h`:
+```bash
+uv run modal run modal_app.py::export --seed 0
+```
+
+### 7. cache audio for competitor evaluation
+regenerates the 5 held-out evaluation sets (clean, windy, dns synthetic, speech+noise, pure noise) with raw audio preserved so all models score identical windows:
+```bash
+uv run modal run modal_app.py::cache_eval_audio
+```
+
+### 8. run all evaluations in parallel
+```bash
+# terminal 1: held-out evaluation on pulsevad models (teacher, 2.1k fp32, 2.1k int8)
+uv run modal run modal_app.py::eval_heldout --seed 0
+
+# terminal 2: score silero-vad v5 on the exact same 200 ms windows
+uv run modal run modal_app.py::eval_competitors
+
+# terminal 3: score marblenet on the exact same windows (nemo container)
+uv run modal run modal_app.py::eval_marblenet
+
+# terminal 4: multilingual benchmark across 10 global & indian languages (fleurs)
+uv run modal run modal_app.py::eval_multilingual
+```
+
+### 9. assemble cross-vad benchmark comparison
+aggregates measured rows and cited competitor numbers into the final comparison artifacts:
+```bash
+uv run modal run modal_app.py::build_comparison --seed 0
+```
+
+### 10. download local artifacts
+pull the final reports, c header, onnx graphs, and benchmark tables to your local workspace:
+```bash
+uv run modal volume get --force pulsevad-data runs/pruned_seed_0/comparison.md ./data/runs/pruned_seed_0/comparison.md
+uv run modal volume get --force pulsevad-data runs/pruned_seed_0/comparison.json ./data/runs/pruned_seed_0/comparison.json
+uv run modal volume get --force pulsevad-data runs/pruned_seed_0/heldout_report.json ./data/runs/pruned_seed_0/heldout_report.json
+uv run modal volume get --force pulsevad-data runs/pruned_seed_0/competitor_report.json ./data/runs/pruned_seed_0/competitor_report.json
+uv run modal volume get --force pulsevad-data runs/pruned_seed_0/multilingual_report.json ./data/runs/pruned_seed_0/multilingual_report.json
+uv run modal volume get --force pulsevad-data runs/pruned_seed_0/pulsevad_weights.h ./data/runs/pruned_seed_0/pulsevad_weights.h
+uv run modal volume get --force pulsevad-data runs/pruned_seed_0/pulsevad_2.1k.onnx ./data/runs/pruned_seed_0/pulsevad_2.1k.onnx
+uv run modal volume get --force pulsevad-data runs/pruned_seed_0/pulsevad_2.1k_int8.onnx ./data/runs/pruned_seed_0/pulsevad_2.1k_int8.onnx
+```
+
+### 11. generate visual benchmark plots
+```bash
+uv run python scripts/plot_comparison.py
+```
+generates:
+- `data/runs/pruned_seed_0/comparison_graph.png`
+- `data/runs/pruned_seed_0/multilingual_graph.png`