Granite Speech 3.2 8B

What is Granite Speech 3.2 8B?

Granite Speech 3.2 8B is an open-weight speech-language model provided by IBM through the IBM Granite organization on Hugging Face. Its canonical repository identifier is ibm-granite/granite-speech-3.2-8b. The model was released on April 2, 2025, under the Apache 2.0 license.

Its main job is to turn English speech into text. It can also translate English speech into selected target languages, including French, Spanish, Italian, German, Portuguese, Japanese, and Mandarin. The output is text: the model does not synthesize spoken audio, create images, or generate video.

Unlike a hosted speech API with a published per-minute price, Granite Speech 3.2 8B is distributed for local or self-managed deployment. That makes it relevant to organizations that want control over model files, processing infrastructure, and data handling, but it also means that users must provide the hardware, software environment, and operational support.

What is the model designed to do?

The model is optimized for speech inputs rather than ordinary text-only language tasks. Its documented applications include:

• Automatic speech recognition for English recordings
• English speech-to-text transcription
• Translation from English speech into supported European and Asian languages
• Self-hosted enterprise workflows that process recorded or streamed audio outside a hosted vendor API

A typical workflow might involve loading a 16 kHz mono recording, asking the model to transcribe it, and then optionally initiating a second language-processing step to translate or further process the resulting text. IBM recommends using its general Granite language models for text-only tasks because Granite Speech is specialized for audio-based input.

Inputs, outputs, and supported modalities

Granite Speech 3.2 8B accepts audio together with text prompts. The official example documents mono audio sampled at 16 kHz. English is the supported speech language identified in the model documentation, while translation targets include French, Spanish, Italian, German, Portuguese, Japanese, and Mandarin.

The model should therefore be evaluated as a speech-recognition and speech-translation system, not as a general multimodal assistant. The supplied research does not verify tool calling, function calling, JSON mode, structured output, web search, or other agent features for this exact model.

How the two-pass architecture works

Granite Speech 3.2 8B uses a two-pass design. In the first pass, a speech encoder interprets the audio and produces a transcription. A second pass can then be explicitly initiated to process the resulting text with the underlying Granite language model.

This separation is important for application design. A developer should not assume that every request automatically performs speech recognition, translation, and broader language reasoning as one indivisible operation. The application may need to manage the transcription step and any subsequent language-processing step separately.

Conformer speech encoder

The speech encoder contains 10 Conformer blocks. Conformer is an architecture designed to combine convolutional processing, which helps capture local patterns in speech, with attention mechanisms, which help model relationships across a longer audio sequence. The documented encoder was trained with Connectionist Temporal Classification on character-level targets.

The published configuration includes 160 input features derived from 80 log-mel features, a hidden dimension of 1,024, eight attention heads, an attention head size of 128, a convolution kernel size of 15, and an output dimension of 42.

Granite language-model component

The speech encoder connects to Granite 3.2 8B Instruct through a speech projector and temporal downsampler. In simple terms, these components convert the encoder's speech representation into a form that the language model can use.

The underlying language-model component has a documented context length of 128,000 tokens. This is the available context for the Granite 3.2 8B Instruct interaction described by the model documentation. It should not be interpreted as a published maximum duration for audio, because audio length also depends on preprocessing, memory, and deployment conditions. No separate maximum output-token limit was verified for this model.

Deployment, software, and licensing

IBM's documented loading example uses Hugging Face Transformers, AutoProcessor, and AutoModelForSpeechSeq2Seq, together with PyTorch and torchaudio. The example uses trust_remote_code=True and loads an audio file from the Hugging Face Hub. Users deploying the model should review the repository code and execution requirements before enabling remote code.

The model is distributed as BF16 Safetensors files and contains approximately 8 billion parameters. BF16 is a reduced-precision numerical format commonly used to lower memory requirements while retaining a useful range for neural-network computation. Actual hardware requirements, transcription speed, and total operating cost will depend on the selected device, audio volume, batching, and serving setup; the supplied research does not provide a verified speed benchmark or hardware performance table.

The Apache 2.0 license permits broad use subject to its terms. Open weights do not, however, remove the need to assess the legal and privacy implications of the audio being processed. Organizations should separately review consent, retention, access control, and any requirements applying to sensitive recordings.

Main strengths and trade-offs

The clearest strength of Granite Speech 3.2 8B is that it combines speech recognition and selected speech translation in a downloadable model rather than requiring a proprietary hosted endpoint. Teams can inspect and operate the model within their own environment, which may be useful when audio cannot be sent to an external service.

Its Apache 2.0 licensing, Transformers-based deployment path, and approximately 8-billion-parameter size also make it a practical candidate for experimentation and controlled enterprise deployment. The Conformer encoder and Granite language-model component provide a defined architecture for users who need more control than a black-box transcription API offers.

Those benefits come with trade-offs. An 8-billion-parameter model is more demanding to operate than a small speech recognizer, and IBM does not publish hosted per-token or per-minute pricing for this exact open-weight model. Users must account for compute, storage, engineering, monitoring, and maintenance costs. The two-pass workflow may also require more application orchestration than a single-purpose transcription endpoint.

There is no verified provider-published reasoning score, coding score, or general-purpose quality score in the supplied research. Reasoning and coding are not the model's intended strengths, and it should not be selected on the assumption that its 128,000-token language-model context makes it a general reasoning system.

Limitations and current positioning

Granite Speech 3.2 8B is specialized. It is not designed for image understanding, video analysis, speech synthesis, music generation, or ordinary text-only assistant workloads. It also does not have a verified hosted IBM token API, published per-token price, or documented support for common agent features such as function calling and web search.

IBM has published Granite Speech 3.3 8B as a newer Granite Speech version. The 3.2 model remains independently available, but teams starting a new project should compare the current model card and behavior of the newer release before committing to 3.2. The existence of a newer sibling does not by itself establish that 3.3 is faster, more accurate, or cheaper; those comparisons require appropriate testing.

The model documentation identifies April 2024 as the knowledge cutoff in the prompt template for the underlying language-model interaction. This is separate from the model's April 2025 release date. The model card describes training with public automatic-speech-recognition and speech-translation data, including Common Voice, Multilingual LibriSpeech, LibriSpeech, VoxPopuli, AMI, YODAS, CREMA-D, and MELD-related data.

When to choose Granite Speech 3.2 8B

Choose Granite Speech 3.2 8B when the central requirement is English speech recognition or English-to-language speech translation and your team is prepared to run an open-weight model. It is especially suitable when local processing, Apache 2.0 licensing, model inspection, or control over deployment is more important than the simplicity of a managed speech API.

• Choose it for: self-hosted English transcription, selected English speech-translation workflows, and applications that need audio processing under their own operational controls.
• Consider a smaller speech model for: deployments where low compute use, low latency, or simple transcription is more important than combining transcription with language-model processing.
• Consider a hosted speech service for: teams that want usage-based billing, managed scaling, minimal infrastructure work, or a clearly documented API and service-level offering.
• Consider a general Granite language model for: text-only prompting, coding, or general language tasks that do not require audio input.
• Compare Granite Speech 3.3 8B for: new IBM Granite Speech evaluations, since it is identified as a newer version in the supplied model research.

Overall, Granite Speech 3.2 8B is best understood as a self-hosted speech-processing component with a connected language-model stage. Its value lies in open deployment and English speech workflows, not in general-purpose reasoning, code generation, or media creation.