IBM releases Granite Speech 5.0 2B and forecasting ensemble; China Telecom open-sources intent model

Granite Speech 5.0 2B ships under Apache 2.0 and is positioned for accurate transcription, on-device transcription and speaker-attributed transcripts, MarkTechPost reported. IBM also updated two existing models: the 470-million-parameter Turbo CTC for edge use and the 4.1 2B Plus for diarization. The speech model runs through Transformers or vLLM on GPUs, mlx-audio on Apple Silicon, and llama.cpp GGUF on local CPU or Mac, with IBM recommending Q8_0 quantization. IBM recommends audio segments of about two minutes.

On the Open ASR Leaderboard, IBM's model card charts a 5.68% average English word error rate as of Oct. 21, 2026, second among open-weight models behind Qwen3-ASR-1.7B at 5.40%, according to MarkTechPost. The model scores 0.99% on LibriSpeech Clean and 13.79% on the leaderboard's private conversational set. IBM's blog reports a 5.06% average instead; MarkTechPost described that figure as unofficial, dated Oct. 15, and limited to public short-form sets. On the noisy FFASR leaderboard, the model ranks fourth of 42 at 14.80% average WER, behind Qwen3-ASR-1.7B at 13.41%.

The model has a 16-block Conformer speech encoder trained with CTC, a two-layer window q-former that downsamples acoustic embeddings, and an intermediate checkpoint of granite-4.0-1b-base extended to 16k positions with YaRN. IBM trained it on 390,000 hours of public audio plus synthetic data for Japanese ASR, keyword biasing and translation, taking 40 days on eight H100 GPUs. New features include language ID prompting, keyword lists of up to 2,000 terms, punctuation and capitalization on request, and better handling of noise, reverberation and distant microphones. Keyword biasing lifted F1 on Earnings-22 from 67.2 to 90.1 with correct lists and to 82.8 with mixed correct and wrong terms. For translation, X-to-English CoVoST BLEU reaches 53.6 for Portuguese and 40.1 for German; English-to-Mandarin was not evaluated.

Turbo CTC now transcribes seven languages with native punctuation and capitalization and reaches 19,870 RTFx, up from about 12,600 at launch, IBM said. IBM described the 4.1 2B Plus update as a drop-in weight update that cuts average timestamp shift from 210.86 to 98.54, long-form WER from 13.53 to 10.22, and biased-keyword WER from 11.79 to 6.48.

Separately, IBM described a Granite Timeseries Ensemble recipe in a Hugging Face blog post. The default Probabilistic Ensemble combines four pre-trained models: Granite PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1 and TTM-r3, covering patched Transformer, patched Conformer, state-space and MLP-mixer architectures. The recipe returns median and quantile forecasts without retraining or fitting ensemble weights; it stores configuration and downloads member checkpoints at runtime. A second IQR Ensemble gives more weight to members with narrower forecast intervals. The models are permissively licensed, and the recipe is available through granite-tsfm 0.3.11 or later.

China Telecom's Xingchen General AI Lab released Xing4.0-29B-Intent with 2.753 million supervised fine-tuning examples drawn from 61 public datasets across 19 domains and 52 languages, along with model weights and an inference image, QuantumBit reported. The model combines intent classification with parameter extraction and uses a Unified Schema-driven Semantic Parsing framework to output JSON under predefined field rules. It supports multiple intents, parameters and entities in one request. The lab said the model returns tens of tokens per request and completes results in hundreds of milliseconds, and that it is already used in telecom customer service for ticket routing and in smart-home speakers for spoken command parsing.

In the lab's own evaluation of 6,000 questions drawn from 48 public datasets, Xing4.0-29B-Intent scored 58 overall, compared with 49 for the highest general model tested, Gemini-3.6-Flash. On a separate 5,606-question category-decision test, it scored 70, compared with 56 for Jev and 65 for GPT-6-Luna. The lab said the tests covered single-label, multi-label and negative examples; it acknowledged that results reflect its test data and settings. The code is under Apache 2.0, with weights and data on Hugging Face.