What is Granite 4.0 1B Speech?
Granite 4.0 1B Speech is an open-weight speech-language model provided by IBM. Its primary jobs are automatic speech recognition (ASR), which means converting spoken audio into written text, and automatic speech translation (AST), which means translating speech into another language.
The model accepts audio and text prompts but produces text only. For example, an application can provide a recording with an instruction to transcribe it, or ask the model to translate the spoken content into English. It does not natively generate spoken audio, images, or video.
IBM released the model on March 6, 2026, under the Apache 2.0 license. The canonical repository is ibm-granite/granite-4.0-1b-speech. Because the weights are available for local deployment, users can run the model through supported open-source serving and inference frameworks rather than relying on a metered IBM-hosted endpoint.
Supported languages and speech tasks
Granite 4.0 1B Speech supports speech input in six primary languages:
- English
- French
- German
- Spanish
- Portuguese
- Japanese
The model supports transcription for these speech languages and speech translation between the supported languages and English. The model card also identifies English-to-Italian and English-to-Mandarin translation use cases. Those additional translation directions should not be interpreted as broad native speech recognition support for Italian or Mandarin; the documented language coverage depends on the specific task and direction.
A keyword-list biasing feature can help the recognizer favor names, acronyms, product terms, and other domain-specific vocabulary. This is useful when standard speech recognition might otherwise misinterpret specialized terminology, although applications should still test accuracy on their own recordings and vocabulary.
Architecture, inputs, and context
The model adapts Granite 4.0 1B Base for speech by combining three main components: a 16-layer conformer speech encoder, a speech projector with temporal downsampling, and the Granite language-model component. A conformer is an encoder architecture commonly used for speech because it can model both local audio patterns and longer dependencies.
Audio is processed as 16 kHz mono input using log-mel features. The speech representation is downsampled before it reaches the language model, reducing the amount of sequence data that must be processed. In practical terms, applications should prepare compatible mono audio and avoid assuming that arbitrary sample rates or multichannel recordings will behave identically to the documented setup.
The underlying Granite language-model component is documented with a 128,000-token context length. The supplied configuration reports a context value of 131,072 tokens, which is the equivalent binary measurement commonly described as 128K. This is a maximum context specification rather than a guarantee that every long recording can be processed efficiently: audio duration, prompt length, generated text, framework behavior, and available memory all affect usable capacity.
No model-specific maximum output-token limit is documented in the supplied research. Example deployments use different generation settings, including 200 new tokens in a Transformers example and 64 tokens in a vLLM example. These are example limits, not a universal model maximum.
Deployment options and pricing
Granite 4.0 1B Speech can be loaded with Hugging Face Transformers using AutoProcessor and a speech sequence-to-sequence model class. IBM also documents offline and server-style deployment with vLLM, including access through a local OpenAI-compatible chat-completions endpoint. On Apple Silicon systems, MLX Audio support is available through a compatible quantized model.
The deployment choice affects latency, memory use, batching, and operational complexity. Transformers is useful for direct application integration and experimentation. vLLM is more appropriate when a team wants a serving layer for multiple requests. MLX Audio offers a route for compatible Apple Silicon environments. These framework-level options should not be confused with a separate IBM API product.
There is no official IBM-hosted input or output token price documented for this checkpoint. It is distributed as open weights, so the financial cost depends on hardware, electricity, storage, hosting, and the selected inference framework. That can make the model attractive for high-volume or privacy-sensitive workloads, but local deployment transfers infrastructure and maintenance responsibilities to the user.
Main strengths and trade-offs
The most important strength is specialization. Granite 4.0 1B Speech is designed specifically for speech recognition and translation rather than general-purpose text reasoning. Its relatively compact positioning, open license, and support for local inference are useful when an organization needs to keep recordings inside its own environment or avoid usage-based hosted speech charges.
The model also includes practical features for production-oriented transcription. Keyword-list biasing can address recurring terminology, while IBM describes improved speech-encoder training and speculative decoding as ways to improve inference efficiency. The supplied evaluation data reports an average word error rate of 5.52 and an RTFx score of 280.02 on the listed Open ASR leaderboard evaluation. These are provider-reported benchmark results, not guarantees for every accent, microphone, noise level, domain, or language pair.
The trade-off is narrower scope than a general-purpose multimodal or reasoning model. The checkpoint returns text and does not synthesize speech. It is not documented as a tool-calling model, and structured JSON output is not identified as a native capability. Fine-tuning resources are documented, but deployment-specific support for streaming, batching, caching, and structured output depends on the serving framework rather than being guaranteed by the checkpoint itself.
Reasoning, coding, speed, and cost characteristics
Granite 4.0 1B Speech should be evaluated primarily as a speech model. It is not presented as a general-purpose reasoning system or coding assistant. Any text generation it performs is mainly in service of transcription and translation, so it should not be selected for complex analysis, software development, or broad conversational reasoning.
As an editorial comparison rather than an IBM-published specification, the supplied assessment rates its reasoning capability at 2 out of 10 and coding capability at 1 out of 10. The same assessment rates speed at 8 out of 10 and cost at 9 out of 10, reflecting its compact open-weight design and the possibility of local inference. These scores are subjective indicators for comparison, not benchmark facts or promises about a particular hardware setup.
In practice, a smaller specialized model may offer a better speed and cost profile for transcription than a larger general-purpose model, especially when the workload is predictable and runs on private infrastructure. However, actual throughput depends on the accelerator or CPU, quantization, audio duration, concurrency, and serving configuration.
When to choose Granite 4.0 1B Speech
This model is a strong candidate when the central requirement is multilingual speech-to-text or speech translation and the organization wants control over deployment. Suitable examples include:
- Private or locally hosted transcription pipelines
- Customer-support or meeting transcription in the documented languages
- Speech translation to or from English
- Edge applications with limited compute resources
- Domain-specific recognition using keyword biasing
- Research and production testing with open-weight speech models
It is particularly worth considering when data residency, offline operation, or predictable infrastructure spending matters more than access to a fully managed commercial speech API.
When another option may be more appropriate
A managed speech API may be a better fit when a team does not want to operate model servers, optimize hardware, or maintain an inference stack. A different speech model may also be preferable if the application requires language coverage outside the documented set, verified speaker diarization, timestamps, or specialized audio analytics that are not described for this checkpoint.
Granite 4.0 1B Speech is also not the right choice for applications that need spoken responses. It produces text transcription or translation, so a separate text-to-speech system would be required for a voice interface. Likewise, general-purpose reasoning and coding workloads should use a model designed for those tasks rather than treating this speech-specialized checkpoint as a broad assistant.
Bottom line
Granite 4.0 1B Speech is a focused IBM model for multilingual speech recognition and translation. Its Apache 2.0 license, local deployment options, keyword biasing, and compact design make it most compelling for private transcription, translation, and edge workloads. Its limitations are equally important: there is no published hosted token price, no native spoken-audio output, no documented universal output limit, and no evidence that it is intended for general reasoning, coding, or broad audio analytics.