Granite Speech 5.0 Turbo CTC release notes / what's new

Hugging Face and IBM have released Granite Speech 5.0 Turbo CTC, a pair of 470M-parameter English speech recognition models designed for extreme speed and high accuracy. These models can achieve a throughput of over 12,600 RTFx on an NVIDIA H200 GPU, enabling the transcription of more than 3.5 hours of speech in a single second using batched inference.

Model Variants and Licensing

Two versions of the model are available to accommodate different licensing and data requirements:

  • granite-speech-5.0-470m-turboctc: An Apache 2.0 licensed model trained on a smaller dataset.
  • granite-speech-5.0-470m-turboctc-nc: A non-commercial model licensed under CC-BY-NC-SA-4.0, trained on additional data for improved accuracy.

Performance and Benchmarks

Based on unofficial results from the OpenASR Leaderboard public English short-form test sets, both models demonstrate high accuracy and unprecedented throughput:

  • Accuracy: The non-commercial (NC) model achieved an aggregate Word Error Rate (WER) of 4.85%, while the Apache 2.0 model scored 5.00% WER.
  • Throughput: Both models exceeded 12,600 RTFx.
  • Comparative Performance: The NC model generally outperforms the Apache 2.0 model across most test sets, showing a significant advantage on SPGI Speech, though it performs worse on the chunked Earnings22 test.

In official results from the FFASR Leaderboard as of August 25, 2026, the granite-speech-5.0-470m-turboctc-nc model ranked fifth in accuracy and granite-speech-5.0-470m-turboctc ranked ninth, with both being the fastest models on the leaderboard.

Model Architecture

Granite Speech 5.0 Turbo CTC utilizes an encoder-only architecture, departing from previous Granite Speech models that used an acoustic encoder paired with a projector and a Granite LM with LoRA adapters. This design reduces the memory footprint to 470M parameters and increases throughput by over 20x compared to previous versions.

Technical Specifications

  • Structure: A stack of 16 Conformer blocks.
  • Optimization: Uses Connectionist Temporal Classification (CTC) loss during training.
  • Attention Mechanism: Employs chunkwise attention to prevent quadratic scaling with sequence length.
  • Self-Conditioning: Implemented at the output of the 8th block.

Temporal Subsampling

To reduce the token rate from 100 frames per second (log Mel spectrogram front end) to 12.5 tokens per second, the models employ three stages of 2x subsampling:

  1. First Stage: Successive log Mel feature vectors are stacked via a reshape() operation.
  2. Second and Third Stages: Integrated into the first two Conformer blocks using strided convolutions to perform temporal downsampling.

Tokenization

The models operate at a lower token rate than previous encoders. The granite-speech-5.0-470m-turboctc-nc model uses SentencePiece tokenization, while the granite-speech-5.0-470m-turboctc model uses BPE tokenization. Both tokenizers are trained specifically on speech transcripts.

Training Data

Training involved a mix of natural and synthetic datasets. Both models were trained on the following natural datasets:

Dataset Hours Source
MLS 44,600 Multilingual LibriSpeech
YODAS 8,900 ESPnet YODAS
CommonVoice-17 2,500 Mozilla Common Voice 17.0
Librispeech 960 OpenSLR LibriSpeech ASR
VoxPopuli 500 Facebook VoxPopuli
AMI 150 Edinburgh CSTR AMI
Earnings-22 100 ESB Datasets

The non-commercial (NC) model received additional training on GigaSpeech (10,000 hours) and SPGI Speech (4,900 hours).

Additionally, both models were trained on three synthetic datasets:

  1. Multi-speaker concatenation (General): 2,000 hours generated from MLS, YODAS, CommonVoice-17, VoxPopuli, and AMI.
  2. Multi-speaker concatenation (Earnings-22): 500 hours generated from Earnings-22.
  3. Targeted Utterances: 240 hours of numbers, currencies, and addresses generated via gpt-oss-120b or gpt-oss-20b and synthesized using StyleTTS2.

Usage and Implementation

Granite Speech 5.0 Turbo CTC is natively supported in the transformers library. Users can implement the model using AutoModelForCTC and AutoProcessor from the Hugging Face ecosystem.

Sources

Related

  • Dispatch
    Fine-Tuning Wav2Vec2 for English ASR with Hugging Face Transformers

    Hugging Face provides a detailed guide on fine-tuning the Wav2Vec2 pretrained speech model for English Automatic Speech Recognition (ASR) using Connectionist Temporal Classification (CTC) loss.

  • Dispatch
    Optimizing Bark using 🤗 Transformers

    Hugging Face demonstrates how to reduce the memory footprint and improve inference speed of the Bark text-to-speech model using Better Transformer, half-precision (fp16), and CPU offload.

  • Dispatch
    Hugging Face AutoGPTQ and Transformers Integration

    Hugging Face has integrated the AutoGPTQ library into Transformers, enabling the quantization of LLMs to 8, 4, 3, or 2-bit precision to reduce memory requirements with negligible accuracy loss at 4-bit.

  • Project
    microsoft/VibeVoice

    VibeVoice is Microsoft’s open‑source suite of frontier voice AI models, offering a 7‑B long‑form ASR that handles up to 60 minutes of audio in one pass (speaker diarisation, timestamps, hot‑word support) and a 1.5‑B multi‑speaker TTS that can synthesize up to 90 minutes of speech. A lightweight 0.5 B real‑time streaming TTS runs with ~300 ms latency. All models share a low‑frame‑rate continuous speech tokenizer and a next‑token diffusion architecture, enabling efficient long‑sequence processing. The repo provides Hugging Face model links, Colab demos, CPU‑quantised ASR via BitNet, vLLM GPU inference, and fine‑tuning scripts. Intended for research; Microsoft warns about bias, deep‑fake misuse, and advises against production use without further testing.

  • Dispatch
    Hugging Face Open ASR Leaderboard Trends and Insights

    Hugging Face has updated the Open ASR Leaderboard with new multilingual and long-form tracks, revealing that Conformer encoders paired with LLM decoders currently lead in English accuracy while CTC/TDT decoders offer superior throughput.