ryota-komatsu/SylReg-LM-7B-Instruct

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 25, 2026License:cc-by-nc-sa-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The ryota-komatsu/SylReg-LM-7B-Instruct is a Qwen2ForCausalLM-based instruction-tuned language model developed by ryota-komatsu, fine-tuned from SylReg-LM-7B. This model is specifically designed for speech language modeling, enabling the generation of speech from syllabic units. It integrates with a syllabic unit encoder and a unit-to-speech decoder, making it suitable for advanced speech synthesis and processing tasks.

Loading preview...

Model Overview

The ryota-komatsu/SylReg-LM-7B-Instruct is an English-language instruction-tuned model built on the Qwen2ForCausalLM architecture. It is fine-tuned from the ryota-komatsu/SylReg-LM-7B base model and is licensed under CC BY-NC-SA 4.0. This model is part of a larger system designed for advanced speech processing, specifically focusing on speech language modeling from syllabic units.

Key Capabilities

  • Speech Language Modeling: The model excels at generating speech from sequences of syllabic units, acting as a core component in a unit-to-speech synthesis pipeline.
  • Integration with SylReg Components: It is designed to work seamlessly with a SylRegForSyllableDiscovery encoder for converting waveforms into syllabic units and a FlowMatchingWithBigVGan decoder for synthesizing speech from these units.
  • Instruction-Tuned: The model is instruction-tuned, allowing for controlled generation based on provided prompts, particularly those formatted with syllabic unit tokens.

Training Details

The model was trained on a substantial dataset totaling over 129,364 hours of audio and text data, including diverse sources like LibriSpeech, Libriheavy, Emilia-Large, People's Speech, VoxPopuli, TinyStories, and Cosmopedia-v2. Training involved 15,000 steps with a batch size of 2^21 tokens, utilizing bf16 mixed precision and an AdamW optimizer. The training was conducted on 32 NVIDIA H100 GPUs.

Use Cases

This model is particularly well-suited for research and development in:

  • Advanced Speech Synthesis: Generating high-quality speech from abstract syllabic representations.
  • Speech-to-Speech Translation Systems: As a component for re-synthesizing speech after linguistic processing.
  • Voice Conversion and Manipulation: Exploring novel ways to control speech generation at a syllabic level.

Finetunes
1 models