IBM has released three significant updates to its open-source Granite Speech collection, expanding the capabilities of its enterprise speech recognition models. The new Granite Speech 5.0 2B model improves transcription accuracy and robustness across six languages—English, French, German, Spanish, Portuguese, and Japanese—with enhanced handling of noisy audio and reverberant environments. The model supports keyword biasing with up to 2,000 terms, language identification prompting, and formatted output with punctuation and capitalization on demand. The updates also address speed-critical use cases with an improved Granite Speech Turbo CTC, a lightweight 470-million-parameter model now extended to support seven languages. IBM reports the revised Turbo CTC achieves a 1.6x speed increase to approximately 19,870 real-time factor (RTFx) on standard benchmarks while maintaining accuracy, making it suitable for on-device and edge deployment. The company simultaneously improved its Granite Speech 4.1 2B Plus for multi-speaker scenarios, reducing speaker identification errors by 7 percent and improving timestamp accuracy by over 50 percent. All three models are distributed under the Apache 2.0 license and integrated with major inference frameworks including Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon processors. The releases underscore IBM's strategy to provide granular speech AI solutions tailored to different performance and accuracy trade-offs across enterprise applications.