Transcribes your speech to text as fast as you talk

Granite Speech 5.0 TurboCTC converts English audio into text, and its whole design is aimed at doing that quickly on modest hardware. At 470 million parameters it is a fraction of the size of the models that usually top transcription benchmarks, yet it posts competitive word error rates on the Open ASR leaderboard, including its noisy and reverberant speech tests. The speed comes from the architecture: instead of generating the transcript word by word the way most modern ASR models do, it decodes the entire output in a single non-autoregressive pass.

It was trained on roughly 60,000 hours of public and openly licensed English audio, plus synthetic data targeting the things transcription models usually fumble: phone numbers, currencies, addresses, and web URLs. The weights are Apache 2.0, so commercial use is fine.

There is no demo Space, but the local setup is genuinely small: pip install a recent transformers release, load the model with a few lines from the model card, and point it at an audio file. It runs comfortably without a dedicated GPU, which is the point; IBM pitches it for laptops, smartphones, and edge devices. If you have a folder of voice memos or meeting recordings, batch-transcribing it is a realistic first test.

English only, and there is no browser demo, so trying it means running a few lines of Python yourself. The single-pass decoder is what makes it fast, but on the hardest audio the much larger word-by-word models still hold the accuracy crown.

• Transcribe meetings and voice memos locally, with nothing leaving your machine.
• Batch-convert a podcast or interview archive into searchable text.
• Add offline dictation to an app without shipping audio to a cloud API.
• Caption videos in a processing pipeline where per-minute API costs add up.
• Handle noisy real-world recordings that trip up smaller on-device models.

Try the demo

View model page

Read this issue on neuralexpedition.com →