Intended Use: The model is intended to be used in enterprise applications that involve processing of speech inputs. In particular, the model is well-suited for English, French, German, Spanish and Portuguese speech-to-text and speech translations to and from English for the same languages plus English-to-Japanese and English-to-Mandarin. The model can also be used for tasks that involve text-only input since it calls the underlying granite-3.3-2b-instruct when the user specifies a prompt that does not contain audio.

Generation:

Granite Speech model is supported natively in transformers from the main branch. Below is a simple example of how to use the granite-speech-3.3-2b revision 3.3.2 model.

Usage with transformers

First, make sure to install a recent version of transformers:

pip install transformers>=4.52.4 torchaudio peft soundfile

Then run the code:

import torch
import torchaudio
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
from huggingface_hub import hf_hub_download

device = "cuda" if torch.cuda.is_available() else "cpu"

model_name = "ibm-granite/granite-speech-3.3-2b"
processor = AutoProcessor.from_pretrained(model_name)
tokenizer = processor.tokenizer
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_name, device_map=device, torch_dtype=torch.bfloat16
)
# load audio
audio_path = hf_hub_download(repo_id=model_name, filename="10226_10111_000000.wav")
wav, sr = torchaudio.load(audio_path, normalize=True)
assert wav.shape[0] == 1 and sr == 16000  # mono, 16khz

# create text prompt
system_prompt = "Knowledge Cutoff Date: April 2024.\nToday's Date: April 9, 2025.\nYou are Granite, developed by IBM. You are a helpful AI assistant"
user_prompt = "<|audio|>can you transcribe the speech into a written format?"
chat = [
    dict(role="system", content=system_prompt),
    dict(role="user", content=user_prompt),
]
prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)

# run the processor+model
model_inputs = processor(prompt, wav, device=device, return_tensors="pt").to(device)
model_outputs = model.generate(**model_inputs, max_new_tokens=200, do_sample=False, num_beams=1)

# Transformers includes the input IDs in the response.
num_input_tokens = model_inputs["input_ids"].shape[-1]
new_tokens = torch.unsqueeze(model_outputs[0, num_input_tokens:], dim=0)
output_text = tokenizer.batch_decode(
    new_tokens, add_special_tokens=False, skip_special_tokens=True
)
print(f"STT output = {output_text[0].upper()}")

Usage with vLLM

First, make sure to install the latest version of vLLM:

pip install vllm --upgrade

• Code for offline mode:

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams
from vllm.assets.audio import AudioAsset
from vllm.lora.request import LoRARequest

model_id = "ibm-granite/granite-speech-3.3-2b"
tokenizer = AutoTokenizer.from_pretrained(model_id)

def get_prompt(question: str, has_audio: bool):
    """Build the input prompt to send to vLLM."""
    if has_audio:
        question = f"<|audio|>{question}"
    chat = [
        {
            "role": "user",
            "content": question
        }
    ]
    return tokenizer.apply_chat_template(chat, tokenize=False)

# NOTE - you may see warnings about multimodal lora layers being ignored;
# this is okay as the lora in this model is only applied to the LLM.
model = LLM(
    model=model_id,
    enable_lora=True,
    max_lora_rank=64,
    max_model_len=2048, # This may be needed for lower resource devices.
    limit_mm_per_prompt={"audio": 1},
)

### 1. Example with Audio [make sure to use the lora]
question = "can you transcribe the speech into a written format?"
prompt_with_audio = get_prompt(
    question=question,
    has_audio=True,
)
audio = AudioAsset("mary_had_lamb").audio_and_sample_rate

inputs = {
    "prompt": prompt_with_audio,
    "multi_modal_data": {
        "audio": audio,
    }
}

outputs = model.generate(
    inputs,
    sampling_params=SamplingParams(
        temperature=0.2,
        max_tokens=64,
    ),
    lora_request=[LoRARequest("speech", 1, model_id)]
)
print(f"Audio Example - Question: {question}")
print(f"Generated text: {outputs[0].outputs[0].text}")


### 2. Example without Audio [do NOT use the lora]
question = "What is the capital of Brazil?"
prompt = get_prompt(
    question=question,
    has_audio=False,
)

outputs = model.generate(
    {"prompt": prompt},
    sampling_params=SamplingParams(
        temperature=0.2,
        max_tokens=12,
    ),
)
print(f"Text Only Example - Question: {question}")
print(f"Generated text: {outputs[0].outputs[0].text}")

• Code for online mode:

"""
Launch the vLLM server with the following command:

vllm serve ibm-granite/granite-speech-3.3-2b \
    --api-key token-abc123 \
    --max-model-len 2048 \
    --enable-lora  \
    --lora-modules speech=ibm-granite/granite-speech-3.3-2b \
    --max-lora-rank 64
"""

import base64

import requests
from openai import OpenAI

from vllm.assets.audio import AudioAsset

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "token-abc123"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    # defaults to os.environ.get("OPENAI_API_KEY")
    api_key=openai_api_key,
    base_url=openai_api_base,
)

base_model_name = "ibm-granite/granite-speech-3.3-2b"
lora_model_name = "speech"
# Any format supported by librosa is supported
audio_url = AudioAsset("mary_had_lamb").url

# Use base64 encoded audio in the payload
def encode_audio_base64_from_url(audio_url: str) -> str:
    """Encode an audio retrieved from a remote url to base64 format."""
    with requests.get(audio_url) as response:
        response.raise_for_status()
        result = base64.b64encode(response.content).decode('utf-8')
    return result

audio_base64 = encode_audio_base64_from_url(audio_url=audio_url)

### 1. Example with Audio
# NOTE: we pass the name of the lora model (`speech`) here because we have audio.
question = "can you transcribe the speech into a written format?"
chat_completion_with_audio = client.chat.completions.create(
    messages=[{
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": question
            },
            {
                "type": "audio_url",
                "audio_url": {
                    # Any format supported by librosa is supported
                    "url": f"data:audio/ogg;base64,{audio_base64}"
                },
            },
        ],
    }],
    temperature=0.2,
    max_tokens=64,
    model=lora_model_name,
)


print(f"Audio Example - Question: {question}")
print(f"Generated text: {chat_completion_with_audio.choices[0].message.content}")


### 2. Example without Audio
# NOTE: we pass the name of the base model here because we do not have audio.
question = "What is the capital of Brazil?"
chat_completion_with_audio = client.chat.completions.create(
    messages=[{
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": question
            },
        ],
    }],
    temperature=0.2,
    max_tokens=12,
    model=base_model_name,
)

print(f"Text Only Example - Question: {question}")
print(f"Generated text: {chat_completion_with_audio.choices[0].message.content}")

Model Architecture:

The architecture of granite-speech-3.3-2b revision 3.3.2 consists of the following components:

(1) Speech encoder: 16 conformer blocks trained with Connectionist Temporal Classification (CTC) on character-level targets on the subset containing only ASR corpora (see configuration below). In addition, our CTC encoder uses block-attention with 4-seconds audio blocks and self-conditioned CTC from the middle layer.

(2) Speech projector and temporal downsampler (speech-text modality adapter): we use a 2-layer window query transformer (q-former) operating on blocks of 15 1024-dimensional acoustic embeddings coming out of the last conformer block of the speech encoder that get downsampled by a factor of 5 using 3 trainable queries per block and per layer. The total temporal downsampling factor is 10 (2x from the encoder and 5x from the projector) resulting in a 10Hz acoustic embeddings rate for the LLM. The encoder, projector and LoRA adapters were fine-tuned/trained jointly on all the corpora mentioned under Training Data.

(3) Large language model: granite-3.3-2b-instruct with 128k context length (https://huggingface.co/ibm-granite/granite-3.3-2b-instruct).

(4) LoRA adapters: rank=64 applied to the query, value projection matrices

Training Data:

Overall, our training data is largely comprised of two key sources: (1) publicly available datasets (2) Synthetic data created from publicly available datasets specifically targeting the speech translation task. A detailed description of the training datasets can be found in the table below:

Infrastructure: We train Granite Speech using IBM's super computing cluster, Blue Vela, which is outfitted with NVIDIA H100 GPUs. This cluster provides a scalable and efficient infrastructure for training our models over thousands of GPUs. The training of this particular model was completed in 13 days on 32 H100 GPUs.

Ethical Considerations and Limitations:

The use of Large Speech and Language Models can trigger certain risks and ethical considerations. Although our alignment processes include safety considerations, the model may in some cases produce inaccurate, biased, offensive or unwanted responses to user prompts. Additionally, whether smaller models may exhibit increased susceptibility to hallucination in generation scenarios due to their reduced sizes, which could limit their ability to generate coherent and contextually accurate responses, remains uncertain. This aspect is currently an active area of research, and we anticipate more rigorous exploration, comprehension, and mitigations in this domain.

IBM recommends using this model for automatic speech recognition and translation tasks. The model's modular design improves safety by limiting how audio inputs can influence the system. If an unfamiliar or malformed prompt is received, the model simply echoes it with its transcription. This minimizes the risk of adversarial inputs, unlike integrated models that directly interpret audio and may be more exposed to such attacks. Note that more general speech tasks may pose higher inherent risks of triggering unwanted outputs.

To enhance safety, we recommend using granite-speech-3.3-2b alongside Granite Guardian. Granite Guardian is a fine-tuned instruct model designed to detect and flag risks in prompts and responses across key dimensions outlined in the IBM AI Risk Atlas.

Resources

• 📄 Read the full technical report: https://arxiv.org/abs/2505.08699 (covers initial release only)
• 🔧 Notebooks: Finetune on custom data, two-pass spoken question answering
• ⭐️ Learn about the latest updates with Granite: https://www.ibm.com/granite
• 🚀 Get started with tutorials, best practices, and prompt engineering advice: https://www.ibm.com/granite/docs/
• 💡 Learn about the latest Granite learning resources: https://ibm.biz/granite-learning-resources