Voice AI and speech recognition
Trelis Research
Trelis Research builds voice AI systems for developers and enterprises working with overlapping speakers, specialist terminology and noisy recordings. Tiron transcribes multi-speaker meetings and attributes speech to speakers, with access through the Trelis API or locally deployed open weights. Tara handles Hindi transcription and Hindi-English code-switching, writing English in Latin script and Hindi in Devanagari. Its open weights work with whisper-large-v3 tooling.
Based in Dublin, Trelis was founded and is led by Ronan McGovern, previously co-founder and CEO of Sandymount Technologies, an MIT spinout acquired by Alfa Laval in 2020. Its engineering work includes Tiron’s Whisper Large-based transcription model and an open-source harness that joins short audio segments into long meeting transcripts. The harness uses two transcription passes and speaker embeddings to link speakers across segments, including overlapping conversations. Tiron is non-streaming, processing audio in 30-second windows.
Trelis offers custom speech recognition and text-to-speech model development, voice agent pipelines, benchmarking, audio data preparation and deployment support. Its Converse product combines recognition, turn-taking, interruption handling, reasoning, tool use and speech generation in one managed real-time API. As of August 2026, Converse remains in private alpha with a small group of design partners.
2 talks
Newest first
Text-to-Speech Data Preparation and Fine-tuning Workshop - Ronan McGovern
AI Engineer World's Fair 2025 · 34:00
1 speaker at AIE
Affiliations reflect their AIE appearances, not necessarily current employment.
Start here
- MCP Agent Fine-Tuning Workshop - Ronan McGovern
Start here to learn how successful agent runs become a small training dataset for examining Qwen3 fine-tuning and tool-calling performance.
Ronan McGovernAI Engineer World's Fair 2025
- Text-to-Speech Data Preparation and Fine-tuning Workshop - Ronan McGovern
Choose this workshop for the speech-data workflow: preparing YouTube audio with Whisper and running adapter-based training in Google Colab with Unsloth.
Ronan McGovernAI Engineer World's Fair 2025
Messages from the stage
Tool compatibility before agent training
The agent workshop connects MCP servers with a Qwen-compatible model endpoint and translates tool descriptions and Hermes-style JSON/XML calls. This integration work accompanies the collection of multi-turn reasoning and browser-use traces.
Voice cloning versus fine-tuning
The speech workshop distinguishes voice cloning from fine-tuning while explaining CSM-1B's hierarchical, two-transformer audio-token architecture. It compares generated speech before and after training.
Affiliations reflect each recorded session, not necessarily current employment.