IBM Releases Granite Speech 5.0 Turbo CTC: 470M-Parameter ASR Model Hits 12,600x Real-Time Speed

IBM has released Granite Speech 5.0 Turbo CTC, a pair of 470M-parameter English speech recognition models that the company claims achieve over 12,600 RTFx (real-time factor) on an NVIDIA H200 GPU — equivalent to transcribing more than 3.5 hours of audio in one second using batched inference.

The release includes two variants: granite-speech-5.0-470m-turboctc, licensed under Apache 2.0, and granite-speech-5.0-470m-turboctc-nc, a more accurate model trained on additional data but restricted to a CC-BY-NC-SA-4.0 non-commercial license.

Benchmark Results

On public short-form test sets from the OpenASR Leaderboard, IBM reports the Apache 2.0 model scored an aggregate 5.00% word error rate (WER), while the non-commercial variant scored 4.85% WER. IBM notes these are unofficial results computed independently using OpenASR Leaderboard tooling and HF Jobs, pending official leaderboard confirmation. The non-commercial model showed a notable accuracy edge on SPGI Speech but underperformed on the chunked Earnings22 test relative to the Apache 2.0 version.

On the FFASR Leaderboard, which measures far-field speech recognition, IBM states the official rankings as of August 25, 2026 placed the Apache 2.0 model ninth in accuracy and the non-commercial model fifth — with both ranking as the fastest two models on the leaderboard.

Architecture Change

Unlike previous Granite Speech models, which combined an acoustic encoder, a projector, and a Granite LM with LoRA adapters, the 5.0 Turbo CTC models are encoder-only. IBM says this design cuts the parameter count to 470M and delivers over 20x faster throughput than prior Granite Speech releases, at the cost of capabilities tied to the language model component, such as speech translation and keyword biasing.

The architecture uses a stack of 16 Conformer blocks with chunkwise attention and self-conditioning at the 8th block, trained with connectionist temporal classification (CTC) loss — consistent with earlier Granite Speech encoders. The key change is token rate: the new models output 12.5 tokens per second, down from the 50 characters per second of previous encoders, achieved through three stages of 2x subsampling applied to a 100 frames-per-second log Mel spectrogram front end. The non-commercial model uses SentencePiece tokenization; the Apache 2.0 model uses BPE tokenization.

Training Data

Both models train on a shared set of natural speech datasets, including Multilingual LibriSpeech (44,600 hours), YODAS (8,900 hours), Common Voice 17 (2,500 hours), LibriSpeech (960 hours), VoxPopuli (500 hours), AMI (150 hours), and Earnings-22 (100 hours). The non-commercial variant adds GigaSpeech (10,000 hours) and SPGI Speech (4,900 hours). Both models also train on roughly 2,740 hours of synthetic data, including concatenated multi-speaker audio and TTS-generated utterances covering numbers, currencies, and addresses, synthesized with StyleTTS2 using text generated by gpt-oss-120b and gpt-oss-20b.

The models are natively supported in Hugging Face Transformers (currently requiring an install from source ahead of the next release) and can be loaded via AutoModelForCTC and AutoProcessor. IBM has also published a WebGPU demo of streaming recognition, though it currently runs only in Chrome and Edge.

What this means

The encoder-only design trades away speech translation and LM-based features for a dramatic footprint and speed reduction — a clear bet on edge and on-device transcription rather than server-side multitask speech understanding. The 12,600 RTFx figure and sub-5% WER, if confirmed on the official OpenASR and FFASR leaderboards, would place these among the fastest open speech recognition models available, though IBM's own framing of the OpenASR numbers as "unofficial" warrants waiting for leaderboard confirmation before treating them as settled. The dual-license strategy — a slightly less accurate Apache 2.0 model alongside a more accurate but non-commercial one — mirrors a pattern increasingly common among model providers trying to serve both open-source developers and data-hungry accuracy-first use cases from a single training pipeline.