Toucan: Computational Science Tools

• Toucan is a diverse suite of computational tools, datasets, protocols, and frameworks enabling advanced research in fields such as asteroseismology, NLP, and network security.
• Its implementations include a virtual observatory for stellar modelling with tight density–frequency mappings, token-aware language models achieving up to 24× speedup, and modular text-to-speech pipelines with robust homograph handling.
• Toucan further extends to secure CAN-bus communications, machine translation for African languages with significant BLEU gains, and large-scale agentic data generation providing open-access benchmarks and reproducible protocols.

Toucan refers to a diverse set of state-of-the-art tools, datasets, protocols, and frameworks across asteroseismology, natural language processing, speech synthesis, network security, machine translation, and agentic data generation. The term’s recurrence across these disparate areas reflects its association with leading-edge, open-access resources or platforms in computational science and engineering.

1. Virtual Observatory Tool: TOUCAN in Asteroseismology

The TOUCAN Virtual Observatory gateway is a web-based interface for querying, visualizing, and extracting properties from precomputed stellar equilibrium and oscillation model grids, specifically optimized for the asteroseismic study of intermediate-mass, main-sequence stars such as δ Scuti-type pulsators. TOUCAN sits atop a set of VO-compliant databases accessed via S3 protocol and harmonizes heterogeneous model collections by exposing a unified data model with ~17 global, ~35 seismic, and ~44 shell parameters, enabling bulk queries and on-the-fly plotting without dedicated client code (Suárez et al., 2014).

TOUCAN’s core utility is in extracting the large frequency separation, $\Delta\nu$, in the low-frequency (radial order $n = 1 \ldots 8$) domain of main-sequence δ Scuti oscillation spectra. The tool supports filtering by stellar parameters (mass, $T_{\mathrm{eff}}$, log g, metallicity, $\alpha_{\mathrm{MLT}}$, $d_{\mathrm{ov}}$), returns computed oscillation spectra, and enables direct computation and averaging of $\Delta\nu$ over selected radial orders and degrees ($\ell = 0\dots3$). The major theoretical result is a tight power-law relation between large separation and stellar mean density:

$$\frac{\Delta\nu}{\Delta\nu_\odot} \simeq 0.776\left(\frac{\langle\rho\rangle}{\rho_\odot}\right)^{0.46}$$

with $\Delta\nu_\odot=134.8\,\mu$Hz, $\rho_\odot=1.48$ g cm$n = 1 \ldots 8$0. This mapping, valid for stars rotating up to $n = 1 \ldots 8$1 break-up velocity, enables independent direct estimation of average density purely from seismic data, supporting tight constraints on mode identification, internal structure, and precise exoplanet radius estimates.

2. Token-Aware Character-Level Language Modeling: Toucan

Toucan (in NLP) is an augmentation for Hourglass Transformer-based character-level LLMs, introducing a dynamic tokenization mechanism learned end-to-end to produce “token-aware” decoders (Fleshman et al., 2023). Toucan’s architecture comprises: (i) a tokenizer (with a transformer-based boundary predictor and pooling aggregators), (ii) a token-level transformer over pooled token representations, and (iii) a character-level decoder. The key advance is explicit end-of-token (EOT) marker supervision during training, which enables, at inference, blockwise character generation within learned token boundaries, significantly reducing per-character decoding cost versus baseline character models.

Mathematically, boundary detection leverages a Gumbel-Sigmoid relaxation for differentiable segmentations, and a compression-rate regularization allows control over average token length. Quantitatively, Toucan achieves up to $n = 1 \ldots 8$2 speedup in character generation with negligible loss in standard language modeling performance (measured in BPC and BPT) and learns a longer, more semantically coherent token distribution compared to BPE/WordPiece. Optimization of boundary placement for LM task (as opposed to mere frequency) yields tokens often aligning to full morphemes, words, or predictive phrases.

3. Speech Synthesis: IMS ToucanTTS System

The IMS ToucanTTS (text-to-speech) system factors speech generation into modular, pipeline stages: (1) text-to-phoneme, (2) rule-based French homograph disambiguation, (3) non-autoregressive phoneme-to-spectrogram conversion via a Conformer + Glow flow, and (4) waveform synthesis using a multi-discriminator GAN vocoder (Lux et al., 2023). The phoneme encoder incorporates articulatory feature vectors, and the homograph disambiguation leverages cascaded POS-based rules plus heuristics, achieving 84% correctness on disambiguation for nearly 800 French homographs.

The spectrogram generator uses FastSpeech2/FastPitch architectures with duration, pitch, and energy predictors, CTC-guided MAS alignment, and a flow-based Glow postnet for spectrotemporal sharpening. The GAN-based vocoder ensembles BigVGAN, MelGAN, HiFi-GAN, and Avocodo discriminators. The entire system (~46M parameters) attains high intelligibility and robust homograph handling but is rated lower than fully end-to-end models in naturalness as measured by MOS.

4. Network Security Protocol: TOUCAN for CAN

TOUCAN is also the designation for a CAN-bus security protocol engineered for in-vehicle networks, achieving authenticity, integrity, and confidentiality while remaining fully CAN 2.0 and AUTOSAR SecOC profile compliant (Bella et al., 2021). The protocol replaces the 64-bit CAN Data field with a concatenation of a 40-bit payload and 24-bit truncated MAC (Chaskey-128), encrypted using AES-128 in CTR mode. No hardware changes to ECUs are required; key material (MAC and encryption keys) is pre-shared.

On an STM32F407 Discovery board (ARM Cortex-M4, 84 MHz), the total cryptographic processing per frame is under 24 μs against a wire latency of ~128 μs. Security claims are predicated on the strength of AES-128 and Chaskey, with authenticity failure rates bounded by $n = 1 \ldots 8$3, though freshness/replay protection is not provided by default. The reduction in payload size and lack of integrated key management are acknowledged limitations.

5. Many-to-Many Machine Translation: Toucan for African Languages

Toucan refers to a family of machine translation (MT) models, fine-tuned from the “Cheetah” encoder-decoder Transformer backbones (580 M, 1.2 B, and 3.7 B parameters), targeting 156 African language pairs (Elmadany et al., 2024). Using the AfroLingu-MT benchmark (43 African languages plus Arabic, English, French; 586K parallel sentences), Toucan outperforms mT5, mT0, and AfriTeVa-v2 models. On the spBLEU$n = 1 \ldots 8$4 test metric, Toucan-3.7B achieves 23.15 versus mT5-xl’s 15.95, with notable improvements in both high- and extremely low-resource directions.

The evaluation introduces the spBLEU$n = 1 \ldots 8$5 metric, leveraging a SentencePiece tokenizer trained on 1K+ languages. Toucan models show robust scaling gains—each capacity increase yields up to +5 BLEU points—and superior lexical and domain coverage, attributed to pretraining on 517 African and 10 global languages, unlike prior approaches. The suite and benchmark are positioned as state-of-the-art resources for advancing inclusive NLP and MT for under-resourced regions.

6. Tool-Agentic Data Generation: TOUCAN Dataset for LLM Agents

In the agentic data domain, Toucan designates the largest open-source tool-agentic dataset: 1.5M multi-turn, multi-tool LLM-agent trajectories generated from 495 real-world Model Context Protocol (MCP) servers (Xu et al., 1 Oct 2025). The five-stage pipeline comprises MCP onboarding (with server smoke-testing), tool-use task synthesis via multiple open LLMs, model-based and rule-based filtering, and agentic trajectory execution using GPT-OSS-120B, Kimi-K2, and Qwen3-32B orchestrated with two agentic frameworks. Three Extension modules introduce irrelevance, persona-based variants, and multi-turn dialogue splits to increase data and behavioral diversity.

Table: Toucan Datasets, Models, and Domains

Subfield	Toucan Resource	Core Functionality/Claim
Asteroseismology	VO Gateway TOUCAN	Stellar model querying/Δν-ρ
NLP (LM)	Toucan Model	Token-aware char-LM
Speech Synthesis	IMS ToucanTTS	Modular TTS pipeline
In-vehicle Security	TOUCAN Protocol	Secure CAN-bus comms
MT for Africa	Toucan (Cheetah fine)	156-pair MT/deep scaling
Agentic Data	TOUCAN Dataset	1.5M tool-agentic trjs

Toucan-tuned agentic models (Qwen2.5-32B-Toucan) extend the Pareto frontier on industry-standard leaderboards such as BFCL V3 and MCP-Universe. The dataset exposes LLMs to planning, real-world tool invocation, trajectory completeness/conciseness, and robust handling of parallel/multi-step/irrelevance scenarios—facilitating SFT of compact models that outperform much larger, closed-source models in real-world automated agent tasks.

7. Significance, Impact, and Limitations

The proliferation of the Toucan name across independent, high-profile computational platforms signals both its generality as a brand and its association with advanced, accessible, and high-utility systems optimized for heterogeneity, interoperability, or open-ended extensibility. Each Toucan artifact targets community needs that include—rapid model query and exploration in stellar physics, fast and tokenization-free language modeling, composable/evaluable TTS pipelines, authenticated automotive systems, inclusive MT for low-resource languages, and open-access agentic data at scale.

A recurring limitation is observed in the partial coverage or domain scope: e.g., the asteroseismic TOUCAN grid is bounded to non-rotating or moderately rotating sources, agentic data servers lack private API coverage, and automotive TOUCAN does not secure against replay without higher-level sequence checks. Nonetheless, each instance advances methodological rigor and provides openly reproducible protocols and benchmarks pivotal for their respective communities.