Contents

IBM's Granite Speech release offers separate models for accurate transcription and translation, fast transcription, and speaker-aware transcription. Choose one that fits your use case. The latest version number does not mean a single model is best for every task.

Key Takeaways#

  • granite-speech-5.0-2b supports multilingual transcription, translation, and keyword lists of up to 2,000 words.
  • granite-speech-5.0-470m-turboctc is designed for fast transcription in seven languages, with built-in punctuation and capitalization.
  • The updated granite-speech-4.1-2b-plus improves speaker identification, timestamps, and recognition accuracy for long recordings.
  • All models are available under the Apache 2.0 license and support Transformers, but published benchmarks need to be validated with your own recordings and hardware.

Choose Based on the Output You Need#

For accurate transcription or speech translation, start with granite-speech-5.0-2b. It supports transcription in English, French, German, Spanish, Portuguese, and Japanese. Translation works in both directions between English and the other five languages, plus from English into Italian and Mandarin Chinese.

For high-throughput or on-device transcription, try granite-speech-5.0-470m-turboctc. Its 470-million-parameter encoder supports English, Japanese, Spanish, German, French, Italian, and Portuguese. Output includes punctuation and proper capitalization.

For meetings, interviews, and calls where speaker labels and utterance timing matter, use granite-speech-4.1-2b-plus. Despite its older version number, this updated model is the release's designated option for recordings with multiple speakers.

Control Accuracy and Terminology#

The 2-billion-parameter granite-speech-5.0-2b adds language ID prompts that guide transcription by specifying the input language. You can request punctuation and proper capitalization, and supply up to 2,000 keywords to prioritize personal names, product names, and abbreviations in transcription and translation.

According to IBM, its average English word error rate (WER) was 5.06%, ranking second among the open-weight ASR models compared. This result was unofficial as of October 15, 2026, and covers the public English short-form audio test sets on the Open ASR Leaderboard. It does not establish the same ranking for Japanese, long meetings, or your organization's audio.

Use each setting deliberately.

  • Specify the input language when you know it.
  • Supply terms that appear in the recording, and check both recognition results and unwanted insertions.
  • Request formatted output when readers need punctuation and proper capitalization.
  • Evaluate translation separately from transcription, especially for long conversational recordings.

Measure Speed on Your Own Hardware#

According to IBM, Turbo CTC reached 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 at launch. That is about a 1.6× increase while maintaining reported accuracy. RTFx expresses processing speed relative to audio duration; higher values mean higher throughput.

This optimization compiles a single Conformer block and reuses its graph across the stacked blocks. IBM says this also reduces warm-up time. The revised model also adds training on reverberant audio, which is relevant to recordings made with distant microphones or in rooms with strong echoes.

Treat published throughput as a benchmark result, rather than a guarantee of application response time. Measure startup time, processing speed, and memory usage on your target device. IBM's WebGPU streaming demo lets you try transcription in a browser, but you still need to check whether performance is acceptable on the devices you plan to support.

Check Speakers and Timestamps#

The updated granite-speech-4.1-2b-plus retains its architecture, prompt format, and API. IBM describes it as a drop-in replacement with improved weights, making it a first candidate to try if you already use this model.

The reported improvements address several practical issues.

| Metric | Previous Version | Updated Version | |---|---:|---:| | Long-form audio WER | 13.53 | 10.22 | | Word-level diarization error rate | 4.40 | 4.10 | | Speaker-aligned WER, cpWER | 30.50 | 28.48 | | Average timestamp drift | 210.86 | 98.54 |

Long-form audio WER fell by about 24%, and average timestamp drift dropped by more than half. With a 20-word keyword list, WER for prioritized terms also fell by about 45%, from 11.79 to 6.48. These are relative reductions on IBM's evaluation sets. Timestamp drift measures timing error, not word error rate. The supplied announcement materials do not specify its unit.

Next Steps#

Use the model that fits your use case and run a comparative evaluation on representative samples from your own recordings. Include challenging audio and relevant terminology, then record transcription errors, processing time, and speaker-label or timestamp errors before choosing your deployment setup.

Source: https://huggingface.co/blog/ibm-granite/granite-speech-5

Contents