IBM releases Granite 4.0 1B Speech, a compact multilingual speech model for edge AI

IBM has released Granite 4.0 1B Speech, a compact speech-language model for multilingual speech recognition and bidirectional translation. What matters here is not just another release, but IBM's bet on production scenarios, where memory, latency and inference cost are as critical as benchmark quality.

What changed

Granite 4.0 1B Speech replaces the heavier configurations in the Granite Speech line and focuses on efficiency. According to IBM, the model has half the parameters of granite-speech-3.3-2b, yet it gains improved English ASR accuracy, Japanese speech recognition support, keyword list biasing and faster inference thanks to encoder fine-tuning and speculative decoding. The idea is simple: don't grow the model at any cost, but strip out excess weight without losing the core capabilities teams need in real-world operation.

IBM also highlights its training approach. The model is built on granite-4.0-1b-base, which was fine-tuned for speech tasks through modality alignment. The training mix included open ASR and AST corpora, as well as synthetic datasets for Japanese, keyword-biased ASR and speech translation. For developers this is an important signal: IBM is not building a closed, cloud-only voice stack, but developing an open model that can be adapted to their own pipelines and hardware.

Languages and tasks

Granite 4.0 1B Speech targets enterprise scenarios that need both transcription and speech translation in both directions. The base set of supported input languages includes English, French, German, Spanish, Portuguese and Japanese. For translation, IBM positions the model as a tool for speech-to-text and speech translation to and from English for these languages, and separately lists English-to-Italian and English-to-Mandarin scenarios. That makes the release useful not only for call centers and voice interfaces, but also for internal translation pipelines.

• Speech recognition in English, French, German, Spanish, Portuguese and Japanese
• Bidirectional speech translation for pairs involving English
• Separate English-to-Italian and English-to-Mandarin scenarios
• Keyword list biasing for names, brands and abbreviations
• Operation in scenarios where low latency and limited memory matter

Another practical plus is the Apache 2.0 license. For enterprise teams, it reduces friction during the pilot and legal review stage: the model can be deployed locally, built into your own stack, and you don't have to depend on API-only access at an early stage. In a market where many speech systems are available only as cloud services with commercial restrictions, this format offers more freedom for customization, offline deployment and control over data.

Deployment and metrics

According to the model card, Granite 4.0 1B Speech has already taken first place on the OpenASR leaderboard with an average WER of 5.52 and RTFx of 280.02. In the per-dataset breakdown, IBM shows, for example, 1.42 on LibriSpeech Clean, 2.85 on LibriSpeech Other and 3.10 on Tedlium. For releases like this, that is an important argument: the model is positioned not simply as "small and cheap", but as a compact system that still holds a very strong level on standard public tests.

On the deployment side, IBM has tried to remove unnecessary barriers. The model is supported in transformers 4.52.1+, runs through vLLM, and has a separate path via mlx-audio on Apple Silicon. The reference pipeline uses 16 kHz mono audio, the request is built with the `<|audio|>` prefix, and keyword biasing can be added directly in the prompt. Architecturally, Granite Speech remains a two-pass system: first the model turns audio into text, and then, if needed, a separate language model call processes the transcript. For production this is convenient, because recognition and downstream logic can be scaled and tuned independently.

What this means

IBM is betting on the segment of voice AI where the winner is not the biggest model, but the one that can realistically run on limited resources without losing quality. If Granite 4.0 1B Speech gains a foothold in production deployments, the market will get another strong open-source option for local transcription, speech translation and edge services without heavy cloud dependence.