IBM releases granite speech 5.0 2B (huggingface.co)

IBM has announced the release of Granite Speech 5.0 2B, a compact and efficient multilingual automatic speech recognition (ASR) and bidirectional automatic speech translation (AST) model. Supporting six languages—English, French, German, Spanish, Portuguese, and Japanese—the model was trained on a vast dataset of 390,000 hours of audio, combining both public corpora and synthetic datasets to enhance performance specifically for Japanese ASR. Compared to its predecessor, Granite Speech 4.1 2B, this new version focuses on improving transcription accuracy, particularly for noisy environments, and expands its capabilities through features like input language ID prompting, increasing keyword biasing capability, and enhanced performance in long-form conversational speech translation. The significance of Granite Speech 5.0 2B lies in its robust improvements tailored for enterprise applications, making it ideal for industries requiring high accuracy in speech processing. Its architecture utilizes advanced techniques such as a speech encoder with 16 conformer blocks and a specialized transformer for downsampling, which collectively boost the model's performance on tasks involving punctuation and capitalization correction. With its Apache 2.0 license and compatibility with recent versions of the Transformers library, this model demonstrates IBM's commitment to advancing AI-driven language technologies, paving the way for more effective multilingual communication solutions in various applications.