IBM Announces Granite Speech 5.0 Collection
IBM has expanded its enterprise speech capabilities with three major model updates designed to improve transcription accuracy, translation, speed, and speaker attribution.
Radwan AbdulhadiEditor and technology writer
Overview of the Granite Speech 5.0 Updates
IBM has introduced three major updates to its enterprise speech portfolio, collectively covering the full range of enterprise audio workflows. Detailed insights and community discussions regarding these releases can be found on the Hugging Face Blog. Together, the new models address high-accuracy transcription and translation, ultra-fast transcription at the edge, and multi-speaker recording scenarios where precise speaker attribution is required.
Granite Speech 5.0 2B Model Capabilities
The new granite-speech-5.0-2b model serves as the successor to granite-speech-4.1-2b, maintaining the same parameter count while expanding functionality. It handles transcription across English, French, German, Spanish, Portuguese, and Japanese, alongside translation to and from English in those languages, plus English to Italian and Mandarin. Training data comprised 390,000 hours of public audio alongside synthetic data designed for Japanese ASR, keyword biasing, and translation tasks.
Among its technical features, the model offers improved English short-form benchmark accuracy, better resilience against noise and reverberation, and language ID prompting to steer multilingual transcription. Furthermore, it supports keyword lists of up to 2,000 terms to bias transcription toward product names and acronyms, as well as on-demand punctuation and capitalization. The model is released under an Apache 2.0 license with integration support in Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.
Turbo CTC Performance and Multilingual Expansion
Initially introduced in August as an English-only 470M-parameter encoder focused on speed, the revised granite-speech-5.0-470m-turboctc now features multilingual transcription capabilities. The updated architecture supports English, Japanese, Spanish, German, French, Italian, and Portuguese, while adding native punctuation and capitalization. Training improvements also target difficult acoustic environments, enhancing performance in echoey rooms and far-field recordings.
Performance optimizations, such as compiling a single Conformer block and reusing its graph across the stack, reduced warmup time. This pushed the model to 19,870 RTFx on the Open ASR Leaderboard, representing an approximately 1.6x speed increase over the initial launch version while maintaining equivalent accuracy.
Granite Speech 4.1 2B Plus for Multi-Speaker Audio
Designed for complex enterprise recordings like meetings, support calls, and interviews, the updated granite-speech-4.1-2b-plus retains the architecture, prompt format, and API of its predecessor to serve as a drop-in replacement. This revision improves speaker attribution, reducing the word diarization error rate and speaker-attributed word error rate by 7%.
Additional improvements include cutting the average timestamp shift by more than half, decreasing long-form test set word error rates by 24%, and reducing errors on biased keywords by 45% when using 20-term lists. Users seeking high transcription accuracy can reference the detailed evaluations and architecture specifications provided on each model card.
Sources
- Hugging Face BlogGranite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster
Continue chronologically
Related stories
Hugging Face Transformers Now Supports llama.cpp GGUF Models
Hugging Face has added native support for GGUF model files to its Transformers library, enabling developers to run quantized AI locally.
IBM Granite Switch 4.1 Brings Chat and Speech to One Endpoint
Discover IBM's Granite Switch 4.1, combining chat, general and specialized decisions, and speech-to-text tasks from a single hosted endpoint.
Mistral Large 4 Released on Hugging Face
Mistral has launched Mistral Large 4, an open-weight Mixture-of-Experts multimodal model featuring 1T total parameters and strong cybersecurity capabilities.
Reflection launches Beam, a 501B open AI model
Reflection has released Beam, a 501B-parameter open model for coding and agentic work with a 1M-token context window.
