Open Model WeightsIBM GRANITE · FIELD VERIFIED
Primary sources last checked on 24 Oct 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
FRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-10-23
IBM Granite Speech 5.0 2B is an official Apache-2.0 downloadable speech-language model for automatic speech recognition in English, French, German, Spanish, Portuguese and Japanese and selected speech translation involving English. The primary model card describes a 16-block Conformer front end, a training mix of 390,000 hours of public audio plus synthetic examples and keyword lists containing up to 2,000 terms. IBM's reported short-form English word error rate is a publisher evaluation, not an independent Open Model Weights result. This model is different from the updated 470M Turbo CTC and 4.1 2B Plus speaker attribution checkpoints. We classify the 5.0 2B release as Notable until independent ASR and translation evidence supports Leading status; its repository revision is source-verified.
WHY IT MATTERS
Open weights offer a source-auditable alternative to opaque speech processing for organizations with multilingual audio.
A documented processor and supported language matrix allow realistic evaluations of transcription, translation and terminology handling.
MODEL-SELECTION VIEW
IBM introduced Granite Speech 5.0 2B as a downloadable speech-language checkpoint for automatic speech recognition and selected speech-translation directions. The official repository declares Apache 2.0 and provides usage paths through Transformers, vLLM, llama.cpp and mlx-audio. This checkpoint is not a general-purpose text chatbot or a standalone voice-cloning application; it is a specialized audio-to-text model.
The broader Granite Speech announcement also updates the 470M Turbo CTC model and the 4.1 2B Plus speaker-attribution variant, but those are distinct weight repositories. Diarization and high-throughput encoder-only claims from companion checkpoints must not be attributed to the new 5.0 2B model. Register, evaluate and version each artifact separately.
The official model card lists English, French, German, Spanish, Portuguese and Japanese for multilingual automatic speech recognition. IBM also documents translation involving English and each of those languages, plus English-to-Italian and English-to-Mandarin generation. These explicit directions differ from universal language-pair translation or speech generation.
Language identification prompts, punctuation controls and keyword lists up to 2,000 terms are designed to handle specialist vocabulary and real-world conversation. Keyword biasing can improve named-entity recall but can also cause unwanted insertions. Evaluate the trade-off per target language and domain with both positive cases and near-negative examples.
IBM describes training on approximately 390,000 hours of publicly sourced audio with supplementary synthetic material for Japanese tasks, term biasing and translation. The model aligns a Granite 4.0 language backbone to audio using a specialized speech encoder instead of depending on a generic image-text processor.
The source details a 16-block Conformer front end with Connectionist Temporal Classification training and an acoustic representation built from 80-dimensional log-Mel features. Those design details support reproducing input preprocessing and segmentation, but they do not guarantee any particular peak memory, streaming throughput or device support.
IBM reports about 5.06 percent average English WER on its selected public short-form recognition test sets, alongside stronger robustness to noise and reverberation. That number remains a vendor result: token normalization, microphones, speaker accents, punctuation and evaluation corpus composition all change the outcome in a production call center.
For long recordings the model card recommends segmenting audio into clips of roughly two minutes or shorter, for example using voice activity detection. Overlapping voices, code switching and poor microphones can still cause omissions. A short English benchmark must never be presented as proof of meeting diarization, six-language parity or reliable recording-wide translation.
The official examples cover Transformers and other supported frameworks, including mlx-audio on Apple Silicon. A public model card establishes a pathway to attempt deployment, not proof that every runtime version loads each artifact correctly. The exact tokenizer, processor and speech sampling rate should be pinned with the weight files.
Measure the duration of audio processing, transcription and translation separately across representative recording lengths. Report peak RAM/VRAM, real-time factor, batch throughput and warm versus cold startup costs. A nominal 2B model designation cannot determine peak hardware usage, which depends on audio encoder execution, attention caches and serving configuration.
The weight repository declares Apache 2.0, including standard notice obligations. This permissive checkpoint license does not give users the right to collect, retain or share another person's voice. A real deployment requires appropriate consent, role access, data minimization, logging and jurisdiction-specific voice privacy review.
Open Model Weights assigns Notable status because this is a newly available, documented multilingual speech model with usable official weights and concrete inference paths. Leading status would require independently comparable recognition and translation measurements across languages, noisy recordings and hardware setups. Organizations should pilot against their own audio rather than copy a single English WER chart.
IMPLEMENTATION & EVALUATION
Four-step exact-revision ASR and AST evaluation using consented audio and measured runtime; not independently executed by Open Model Weights.
Freeze repository SHA, weight manifest, audio preprocessing versions, prompts and Apache-2.0 evidence.
Deterministic speech-to-text and speech-translation test cases with a versioned manifest.
Resampling or dependency changes distort evaluation outputs.
Sample all six recognition languages, accents, noise, jargon and microphone conditions.
Auditable references and WER/term recall per language with failure examples.
A short English dataset hides low-resource and domain failures.
Test only declared translation pairs with and without terminology biasing and language-ID prompts.
Translation metrics plus keyword recovery and false insertion scores.
A high bias setting hallucinates requested product names.
Benchmark Transformers and any intended serving runtime with recorded retention and access policies.
Peak memory, real-time factor, latency, permissions and rollback acceptance criteria.
Weight license or advertised model size is mistaken for runtime or privacy proof.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
IBM makes official weights available with an Apache-2.0 declaration, processor files and usage examples; exact SHA and artifact accessibility are verified separately during registry intake.
The published ASR scope is English, French, German, Spanish, Portuguese and Japanese; accuracy depends on recording quality, regional speech and terminology.
Speaker attribution improvements apply to the separate Granite Speech 4.1 2B Plus model, not automatically to the Granite Speech 5.0 2B checkpoint.
No. IBM documents specific speech-translation directions involving English, with additional English-to-Italian and English-to-Mandarin options.
No. IBM publishes results on particular short-form English datasets; Open Model Weights has not executed the audio model or measured its accuracy.
No. Weight licensing and data protection are separate: consent, retention, access controls and voice-data rules must be evaluated for each application.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Granite Speech 5.0 2B as a 2.31321B-parameter speech audio to transcription and speech translation text open-weight model in the Granite Speech 5.0 family, with IBM Granite listed as the developer. The checked source repository provides 2 recognized weight artifacts, format evidence for Safetensors and a verified context window of 16,384 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares apache-2.0. Commercial use allowed under apache-2.0 conditions. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
The checked sources declare ibm-granite/granite-4.0-1b-base as lineage evidence. Training recipe status: training-described-in-model-card; data disclosure status: described-in-model-card.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
Commercial use allowed under apache-2.0 conditions
This section summarizes the checked repository evidence. It is not legal advice.
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
01 · WEIGHTS
02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
MODELS FOR AI AGENTS · ROLE EVIDENCE
Specialized model. A supporting component. This checkpoint does not, by category alone, serve as the planning and tool-execution controller.
The curated notable tier is not an agent-readiness rating. Model-specific tool-call success, long-horizon task accuracy and compatible runtime configurations are not independently established here.
Source-linked license: apache-2.0 (allowed commercial-use classification; see original terms).
Context: 16384 configured tokens.
BF16 weights: 4.63 estimated decimal GB, not runtime VRAM.
Compare agent model candidates → · Agent evaluation method → · Inspect original model repository ↗
These records are separately sourced. They do not replace independent model-plus-runtime tests.
No model-specific agent capability evaluation is recorded for this checkpoint. Missing evidence is unknown, not a failed task.
Read evidence details and benchmarks → · Use the agent model selector →