IBM releases Granite Speech 5.0 with faster transcription and better speaker tracking

IBM has released three updates to its Granite Speech collection, introducing a new speech-language model and revising two existing ones to improve accuracy, speed, and speaker attribution. Published on October 21, 2026, these open-weight models aim to cover the full spectrum of enterprise speech tasks, from high-fidelity transcription to real-time edge processing.

What happened

The centerpiece of this release is granite-speech-5.0-2b, a new model that replaces the previous 4.1 version at the same parameter count. It supports transcription in English, French, German, Spanish, Portuguese, and Japanese, and offers translation capabilities between these languages and English, as well as English to Italian and Mandarin. The model was trained on 390,000 hours of public audio, supplemented by synthetic data designed to improve Japanese automatic speech recognition (ASR), keyword biasing, and translation performance.

Alongside the new 2B model, IBM updated granite-speech-5.0-470m-turboctc. Originally an English-only encoder focused on speed, this revision adds multilingual support for seven languages and native punctuation and capitalization. It also features improved robustness against reverberant audio and echoey rooms. Through technical optimizations that compile a single Conformer block and reuse its graph, the model achieved a significant speed boost, reaching 19,870 RTFx on the Open ASR Leaderboard, up from approximately 12,600 RTFx at launch.

The third update targets multi-speaker scenarios with granite-speech-4.1-2b-plus. This model retains the architecture and API of its predecessor but delivers improved weights for better speaker attribution and timestamp precision. It is designed as a drop-in replacement for enterprises dealing with meetings, support calls, and interviews where identifying who spoke and when is critical.

How it works

The new granite-speech-5.0-2b model introduces language ID prompting, allowing users to specify the input language to steer multilingual transcription. It also supports keyword biasing lists of up to 2,000 terms, enabling the model to prioritize specific product names, people, or acronyms during transcription and translation. Users can request formatted output with punctuation and capitalization, achieving capitalization F1 scores between 89.7 and 99.5 across test sets.

For the Turbo CTC model, the performance gain comes from compiling a single Conformer block and reusing its computational graph across the stack. This reduces warmup time and increases throughput without sacrificing accuracy. The model now outputs transcripts with native punctuation and capitalization, eliminating the need for separate post-processing steps.

The granite-speech-4.1-2b-plus model improves speaker attribution by reducing the word diarization error rate (WDER) by 7% and the speaker-attributed word error rate (cpWER) by 7%. It also cuts the average timestamp shift (AAS) by more than half for all speech, providing tighter timing for captions and search indexing. Additionally, it reduces errors on biased keywords by 45% when using 20-term keyword lists.

Key details

• granite-speech-5.0-2b ranks second among top open-weight ASR models on average English word error rate (WER) at 5.06%.
• The Turbo CTC model is now 1.6x faster than its launch version, reaching 19,870 RTFx.
• granite-speech-4.1-2b-plus reduces average timestamp shift from 210.86 to 98.54 for all speech.
• All three models are released under the Apache 2.0 license.
• Native support is available in Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.
• The 2B model supports keyword biasing lists of up to 2,000 terms.

Why it matters

For software engineers building transcription features, these updates offer more control and flexibility. The ability to bias keywords allows applications to handle domain-specific terminology accurately, which is crucial for industries like healthcare, legal, and finance. The improved robustness to noise and reverb means these models can perform better in real-world environments, such as conference rooms or remote calls, without requiring extensive audio preprocessing.

The speed improvements in the Turbo CTC model make real-time, on-device transcription more viable. This is particularly important for applications running on edge devices or in browsers, where latency and resource constraints are significant factors. The native support for punctuation and capitalization simplifies the pipeline, reducing the need for additional models or services to format the output.

Improved speaker attribution and timestamp precision in the 4.1 Plus model enhance the usability of transcripts for meeting summaries and search. Accurate timestamps allow users to jump to specific parts of a recording, while correct speaker labeling makes it easier to follow conversations. These improvements reduce the manual effort required to clean and organize transcribed data.

What you can do

• Evaluate granite-speech-5.0-2b for high-accuracy multilingual transcription and translation tasks.
• Use granite-speech-5.0-470m-turboctc for high-throughput or on-device transcription needs, leveraging its WebGPU streaming demo.
• Deploy granite-speech-4.1-2b-plus for applications involving meetings, calls, or multi-speaker recordings where speaker identification is key.
• Implement keyword biasing in the 2B model to improve accuracy for domain-specific terms and names.
• Test the models with noisy or reverberant audio to assess their robustness in your specific use case.
• Consider pairing these models with Granite Guardian for production deployments that require additional risk detection.