Granite Speech 5.0 Release: Enhanced Transcription, Translation, and Speaker Attribution

The Granite Speech 5.0 collection delivers three major updates to IBM's speech models, covering the full spectrum of enterprise speech needs: high-accuracy multilingual transcription and translation, extremely fast on-device transcription, and detailed multi-speaker recording analysis.

Granite Speech 5.0 2B: High Accuracy and Multilingual Control

The granite-speech-5.0-2b model is the successor to granite-speech-4.1-2b and is designed for high-accuracy, multilingual transcription and translation. It supports transcription of English, French, German, Spanish, Portuguese, and Japanese, and translation to and from English in those languages, plus English to Italian and Mandarin.

This model was trained on 390,000 hours of public audio, supplemented with synthetic data specifically engineered for Japanese ASR, keyword biasing, and translation.

Key improvements over the previous version include:

  • Enhanced Robustness: The model exhibits higher accuracy on English short-form benchmarks and handles noise, reverberation, and distant-microphone speech more effectively.
  • Multilingual Steering: Users can employ Language ID prompting to explicitly name the input language, thereby steering the multilingual transcription process.
  • Advanced Translation: Translation quality is improved for long-form and conversational speech, moving beyond simple short-read sentence translation.
  • Keyword Biasing: The model supports keyword lists of up to 2,000 terms, allowing transcription and translation to be biased toward specific product names, people, or acronyms across all supported languages.
  • Formatted Output: Punctuation and capitalization can be requested via the prompt, achieving capitalization F1 ranges from 89.7 to 99.5 across test sets.

Unofficially, granite-speech-5.0-2b ranks #2 among top open-weight ASR models on average English WER, achieving a score of 5.06% on the Open ASR Leaderboard’s public English short-form test sets.

Granite Speech 5.0 Turbo CTC: Speed and Edge Deployment

The granite-speech-5.0-470m-turboctc model maintains its small 470M-parameter footprint while significantly enhancing its capabilities for high-throughput and on-device transcription.

This revision addresses key user requests by adding three major features:

  • Multilingual Support: The model now supports transcription in English, Japanese, Spanish, German, French, Italian, and Portuguese.
  • Native Formatting: Transcripts are generated with native punctuation and capitalization, eliminating the need for separate post-processing steps.
  • Acoustic Robustness: Training explicitly targets reverberant audio, making Turbo CTC more robust when processing recordings from echoey rooms or far-field sources.

In terms of performance, the model is faster, achieving 19,870 RTFx on the Open ASR Leaderboard, representing an improvement of approximately 1.6x over its initial launch performance of 12,600 RTFx, while maintaining the same accuracy level.

Granite Speech 4.1 2B Plus: Multi-Speaker and Long-Form Analysis

The updated granite-speech-4.1-2b-plus is specifically tuned for complex enterprise recordings such as meetings, support calls, and interviews involving multiple speakers over extended periods. This model is a drop-in replacement for the current version, retaining the original architecture, prompt format, and API while benefiting from improved weights.

This revision delivers significant performance gains across four critical areas:

  • Speaker Attribution: The model improves the accuracy of labeling who is speaking. The Word Diarization Error Rate (WDER) drops by 7%, and the speaker-attributed WER (cpWER) drops by 7% overall.
  • Timestamping: The Average Timestamp Shift (AAS) is reduced by more than half, resulting in tighter timing for captions, search, and review. For short speech, AAS dropped from 29.27 to 26.20; for all speech, it dropped from 210.86 to 98.54.
  • Long-Speech Performance: WER on long-form test sets decreases by 24%, improving from 13.53 to 10.22.
  • Keyword Biasing: Errors related to biased keywords fall by 45% when using 20-term keyword lists, improving from 11.79 to 6.48.

Model Selection and Deployment

The choice of model depends on the primary use case:

  • For high transcription accuracy, translation, or keyword biasing: Use granite-speech-5.0-2b.
  • For raw speed or on-device transcription: Use granite-speech-5.0-470m-turboctc.
  • For meetings or multi-speaker calls: Use granite-speech-4.1-2b-plus.

All models are released under an Apache 2.0 license and offer native support in frameworks including Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon. For production deployments requiring additional risk detection, these models can be paired with Granite Guardian.

Sources