Granite Speech 5.0 2B ranks #2 on Open ASR Leaderboard

IBM has released Granite Speech 5.0 2B, an open-weight speech-language model built for multilingual automatic speech recognition and translation. The model provides teams with a high-accuracy transcription tool they can run on their own infrastructure.

According to IBM’s model card, Granite Speech 5.0 2B ranks second on the Open ASR Leaderboard among open-weight models, achieving a 5.68% average English word error rate across nine public and two private test sets as of October 21, 2026. Only Qwen3-ASR-1.7B scores higher at 5.40%. Granite Speech 5.0 2B supports transcription and translation for six languages: English, French, German, Spanish, Portuguese, and Japanese.

What changed

IBM released Granite Speech 5.0 2B under Apache 2.0 terms, making it available for commercial and private use. The company trained the model on 390,000 hours of public audio data plus synthetic datasets tailored for Japanese ASR, keyword-biased ASR, and speech translation. Training took 40 days on eight H100 GPUs. The model fine-tunes an intermediate checkpoint of IBM’s Granite 4.0 1B base model, which IBM extended to handle 16,000-token positions using YaRN positional extension.

The architecture pairs a speech encoder with 16 Conformer blocks trained via CTC loss with a language model. A two-layer window q-former projects and downsamples acoustic embeddings to 10 Hz, reducing total audio encoding by 10 times. IBM recommends audio segments up to about two minutes in length.

Compared to Granite Speech 4.1 2B, the new version adds language ID prompting, where naming the input language in the prompt steers transcription toward the correct language. It also supports keyword lists containing up to 2,000 terms to bias output toward product names, people, and acronyms. Users can now request native punctuation and capitalization, with capitalization F1 scores ranging from 90.16 to 99.51 across test sets. The model also shows better handling of noise, reverberation, and distant microphone speech.

IBM updated two companion models. Granite Speech 5.0 Turbo CTC now transcribes seven languages with native punctuation and capitalization, reaching 19,870 real-time factor, compared to about 12,600 at launch. The updated Granite Speech 4.1 2B Plus shows average timestamp shift falling from 210.86 milliseconds to 98.54 milliseconds and long-form WER dropping from 13.53% to 10.22%.

Why it matters

This release offers teams a credible open alternative to closed-source transcription services. For organizations with on-device deployment requirements or those preferring to avoid cloud vendors, the #2 leaderboard ranking demonstrates that Granite Speech 5.0 2B can handle production workloads. Open speech models are maturing to a point where enterprise adoption becomes viable.

The keyword biasing feature delivers measurable business value. On the Earnings-22 benchmark, keyword F1 improved from 67.2 to 90.1 when correct keyword lists were provided. Even with 2,000-term lists mixing correct and incorrect keywords, the model still reached 82.8 F1. This matters for financial transcription, call centers, and customer support workflows where domain terminology accuracy is critical.

The multilingual support expands the addressable use case. While English achieves 0.99% WER on LibriSpeech Clean, the model reaches 2.60% on Portuguese CommonVoice and 3.50% on Spanish. Japanese performs weakest at 10.74%, giving teams visibility into where the model excels and where it may need human review. On the private conversational set, performance drops to 13.79% WER, suggesting real-world variability depending on audio characteristics.

The model’s ability to run on Apple Silicon via mlx-audio and on local CPUs via llama.cpp addresses privacy-focused deployments. Teams building customer-facing applications can now keep audio processing on-device rather than transmitting it to cloud endpoints.

What to test

Before adopting Granite Speech 5.0 2B in production, teams should verify several claims:

Leaderboard gap: Test whether the 0.28 percentage point gap versus Qwen3-ASR-1.7B (5.40% WER) matters for your use case. Run sample audio through both models on your domain to measure real performance differences.

Noisy audio handling: The model claims improved robustness to noise and reverberation. Benchmark it against your actual recordings including call center audio, street-level interviews, or conference calls, rather than relying on public benchmark results alone. The model ranks fourth on the FFASR noisy speech leaderboard at 14.80% WER, while Qwen3-ASR-1.7B reaches 13.41%, a meaningful gap.

Language performance variance: Japanese reaches 10.74% WER versus Portuguese at 2.60%. If you support multiple languages, test each one independently. The worst-case scenario on the private conversational set shows 13.79% WER, indicating substantial variation based on audio domain.

Keyword list scaling: The claim of up to 2,000-term keyword lists is new. Test whether performance remains stable as you approach the upper limit and whether keyword order or composition affects results.

Quantization impact: IBM recommends Q8_0 quantization for GGUF inference. The company notes that lower-bit quantizations lose accuracy on audio longer than one minute. Measure this trade-off against your latency requirements on realistic audio lengths.

Timestamp accuracy: The updated 4.1 2B Plus shows average timestamp shift falling from 210.86 to 98.54 milliseconds. If your application relies on word-level timing for synchronization or editing workflows, validate this independently on your data.

Common questions

How does Granite Speech 5.0 2B compare to Whisper?

Granite Speech 5.0 2B ranks second on the Open ASR Leaderboard at 5.68% average WER, indicating stronger English transcription accuracy than Whisper Base. The model includes keyword biasing and language ID prompting features that Whisper does not natively support. However, the source does not provide direct benchmarks comparing the two models on identical test sets.

What languages does the model support?

The model transcribes and translates for English, French, German, Spanish, Portuguese, and Japanese. Translation includes bidirectional capability with X-to-English and English-to-X support. Japanese shows the highest error rate of supported languages at 10.74% WER on CommonVoice, while Portuguese and Spanish perform better.

Can I run this on my laptop or local machine?

Yes. The model runs on GPU via Transformers or vLLM, on Apple Silicon via mlx-audio, and on local CPU or Mac via llama.cpp. IBM recommends Q8_0 quantization for GGUF files, though lower-bit quantizations degrade accuracy on audio longer than one minute.

How much audio data was used to train this?

The model was trained on 390,000 hours of public audio data plus synthetic datasets tailored for Japanese ASR, keyword-biased ASR, and speech translation. Training took 40 days on eight H100 GPUs.

The conclusion

Granite Speech 5.0 2B represents a meaningful step forward for open speech AI. The #2 leaderboard ranking and keyword biasing capability make it appealing for enterprises, especially those needing on-device inference or domain-specific accuracy improvements. The Apache 2.0 license removes licensing friction for commercial use.

The gap to Qwen3-ASR-1.7B is small enough that choice should depend on your specific language mix, infrastructure constraints, and keyword biasing requirements rather than raw leaderboard position. Teams building multilingual systems should test Japanese performance carefully, as it lags other languages significantly.

The model’s real-world adoption will depend on how organizations apply keyword biasing in financial transcription and customer support workflows. Updates to the FFASR leaderboard and domain-specific deployments will indicate whether Granite Speech 5.0 2B sustains its competitive position.

Related reading

• Granite Switch 4.1: Chat, decisions, speech endpoint
• ElevenLabs vs Murf AI: Text-to-Speech Comparison
• Voxtral 3 voice agent SDK: 50+ languages and emotional speech