IBM Granite Speech 5.0 Adds Improved Transcription, Translation, and Speaker Attribution

IBM has released Granite Speech 5.0, an updated version of its open speech model with notable improvements in transcription accuracy, multilingual translation, and speaker attribution (diarization). The model is designed to run faster than its predecessor while improving word error rates across a range of languages and acoustic conditions. Speaker attribution is a particularly valuable addition for developers building meeting transcription, call analytics, or voice-based agent systems where knowing who said what is essential. Granite Speech 5.0 is available on Hugging Face, making it straightforward to integrate into existing pipelines. For teams evaluating open alternatives to proprietary speech APIs like Whisper or Google Speech-to-Text, this is a strong new benchmark to test against.