Model Card for Luciole-23B-Audio-1.0

• Model Description
  • Bias, Risks, and Limitations
  • Recommendations
• Training details
  • Training data
  • Instruction template
  • Training Procedure
• Evaluation
• Using the model
  • With vLLM
  • With NeMo
  • Example prompts for each task
• Acknowledgements
• Contact

Model Description

Luciole-23B-Audio-1.0 is an audio-capable version of Luciole-23B-Instruct-1.1. The model was developed by LINAGORA and the OpenLLM-France consortium, as part of the OpenLLM France project funded by BPI France under the France 2030 program.

The training of Luciole-23B-Audio-1.0 was conducted on the Jean Zay supercomputer managed by IDRIS, using the NVIDIA NeMo Speech toolkit. The model was trained on various tasks and types of audio, including ASR (Automatic Speech Recognition), AST (Automatic Speech Translation), QA (Question Answering), Sound (Question Answering and Captioning), Music (Question Answering and Captioning) and more.

If you are looking for the best compromise between speed and performance go check the Luciole-8B-Audio-1.0.

Bias, Risks, and Limitations

• Inherits the limitations of the base Luciole-23B-Instruct-1.1 language model: it can struggle with math word problems and is susceptible to hallucination.
• Training sequences were capped at 16,384 tokens (audio and text combined), so longer contexts are untested.

• Language coverage is uneven. French and English make up most of the training mix and are the only languages seen on every task. German, Spanish and Italian were seen in training for ASR, AST and spoken QA. Portuguese, Dutch and Arabic for ASR and AST only.
• Arabic is the weakest language: the frozen Parakeet-TDT-0.6B-v3 encoder was not trained on Arabic, so the model is clearly less reliable there (FLEURS WER ~25%).

Recommendations

• Speech, sound and music understanding (summarization, captioning, sound/music QA, conversation): best in French and English, the only languages the model was trained on for these tasks.
• Transcription (ASR) and translation (AST): supported in French, English, German, Spanish, Italian, Portuguese and Dutch. For production ASR/AST pipelines, a dedicated model may still perform better.
• Spoken question answering: mainly supported in French and English. German, Spanish and Italian are also supported, but with far less training data.
• Other languages and tasks: outside French and English, evaluate the model on your own data before relying on it, especially for a task the model was not trained on in that language.
• Up-to-date or domain-specific knowledge: as with the base LLM, pair the model with a RAG pipeline.

Training details

Training data

Trained on the OpenLLM-France/Luciole-Audio-Training-Dataset, a large multilingual, multi-task collection of audio–text conversations comprising: ASR (Automatic Speech Recognition), AST (Automatic Speech Translation), spoken question answering, summarization, diarization, temporal localization, speaker/gender/age/emotion/language recognition, and music/sound captioning and QA (Question Answering).

Instruction template

Same chat template as the base LLM (inspired by Qwen3), extended with an <|audio|> locator tag: wherever <|audio|> appears in a turn, the corresponding audio segment is encoded and its embeddings replace the tag before the sequence reaches the LLM. Conversations can carry several audio clips and mix audio-only, text-only, and audio+text turns within the same dialogue (see the dataset card for example conversations).

Training Procedure

The model was trained for 100k steps and drew from a weighted, randomly-ordered, sharded mix of that dataset, bucketed by duration into 27 buckets (up to 1,200s / 16,384 audio-equivalent tokens per example) with a matching dynamic batch size per bucket (238 down to 1).

Validation covered CommonVoice ASR (fr/en/ar), Multilingual TEDx speech translation (fr→en), spoken QA (SLUE-SQA-5, VoxPopuli-QA, en/fr), and audio/music captioning (AudioCaps, MusicCaps).

The training code and configs are available in the Luciole-Audio-Training repository. Training used an updated version of NeMo/Speech 3.1 which can be found in our NeMo/Speech fork.

Evaluation

To evaluate our models and compare them with other open-weight Audio LLMs, we used our fork of AudioBench. We used a combination of datasets from AudioBench and datasets available in our HF audio evaluation repository. You can find the full details and results on our AudioBench leaderboard.

A small excerpt from the leaderboard, showing the main table across all tasks and languages:

Results for French and English only, as well as for ASR, AST, and QA tasks only:

As you can see, the model OpenLLM-France/Luciole-23B-Audio-1.0 performs well in French and English compared to other models.

For more detailed results in French ASR you can take a look at our 🏆 French ASR Leaderboard where we use various types of datasets to evaluate the open source models.

Using the model

With vLLM

The exported checkpoint's config.json declares "model_type": "nemo_speechlm" and "architectures": ["NeMoSpeechLMForConditionalGeneration"]. Both are registered with vLLM by the SALM plugin that ships inside nemo-toolkit (nemo.collections.speechlm2.vllm.salm) via the vllm.general_plugins entry point:

[project.entry-points."vllm.general_plugins"]
nemo_speechlm = "nemo.collections.speechlm2.vllm.salm:register"

vLLM auto-discovers this plugin at startup as soon as nemo-toolkit and vllm are installed in the same environment. The plugin merges the LoRA adapters into the LLM backbone on load and runs the frozen Parakeet encoder + connector to turn each <|audio|> tag into the right number of audio-embedding slots before generation.

Install using uv

uv venv .venv --python 3.12
source .venv/bin/activate

uv pip install "nemo-toolkit[speechlm2,tts] @ git+https://github.com/linagora-labs/NeMo.git@luciole-1b-vllm-nemotron-fix"
uv pip install vllm==0.28.0
uv pip install "numpy<=2.4"

Serve

vllm serve OpenLLM-France/Luciole-23B-Audio-1.0 \
    --max-model-len 16384

Query (OpenAI-compatible chat API, audio sent as base64):

import base64
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

with open("sample.wav", "rb") as f:   # 16 kHz mono
    audio_b64 = base64.b64encode(f.read()).decode()

response = client.chat.completions.create(
    model="OpenLLM-France/Luciole-23B-Audio-1.0",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Transcris cet audio en français : <|audio|>"},
                {"type": "input_audio", "input_audio": {"data": audio_b64, "format": "wav"}},
            ],
        }
    ],
)
print(response.choices[0].message.content)

The <|audio|> placeholder is the same audio_locator_tag used during training: put it in the text wherever the audio should be attended to; the plugin expands it automatically. Audio must be 16 kHz mono; the encoder supports chunked processing for long-form audio (well beyond the durations seen in training).

With NeMo

The checkpoint can also be loaded directly with NeMo's SALM class (nemo.collections.speechlm2.models.salm).

Install

uv venv .venv --python 3.12
source .venv/bin/activate

uv pip install "nemo-toolkit[speechlm2,tts] @ git+https://github.com/linagora-labs/NeMo.git@luciole-1b-vllm-nemotron-fix"

Load and generate:

import torch
from nemo.collections.speechlm2 import SALM

model = SALM.from_pretrained("OpenLLM-France/Luciole-23B-Audio-1.0")  # or a local checkpoint directory
model = model.eval().to(torch.bfloat16).to("cuda")

# High-level API: pass the audio file path(s) directly in the prompt, next to the
# `<|audio|>` placeholder; SALM loads and resamples the audio for you.
answer_ids = model.generate(
    prompts=[
        [
            {
                "role": "user",
                "content": f"Transcris cet audio en français : {model.audio_locator_tag}",
                "audio": ["sample.wav"],
            }
        ]
    ],
    max_new_tokens=256,
)
print(model.tokenizer.ids_to_text(answer_ids[0].tolist()))

A prompt can carry several turns and several audio clips (one <|audio|> tag per clip, in order).

Example prompts for each task

• ASR:
  • English: Make a word-for-word transcription in English of the spoken content.
  • French: Fais une transcription mot pour mot en français du contenu parlé.
  • German: Erstelle eine wortwörtliche Transkription des gesprochenen Inhalts auf Deutsch.
  • Spanish: Haz una transcripción palabra por palabra en español del contenido hablado.
  • Italian: Fai una trascrizione parola per parola in italiano del contenuto parlato.
  • Portuguese: Faça uma transcrição palavra por palavra em português do conteúdo falado.
  • Dutch: Maak een woord-voor-woord transcriptie in het Nederlands van de gesproken inhoud.
  • No language given: Can you write the text of this audio file in its source language?
• QA: Listen to the audio and answer the question.
• AST: Translate this file into French.
• Diarization:
  • Provide a time-stamped, speaker-attributed transcript of this audio.
  • Transcris et diarise cet enregistrement, avec les horodatages de chaque tour.

Acknowledgements

Training of Luciole-23B-Audio-1.0 was made possible by computing AI and storage resources by GENCI at IDRIS thanks to the grant 2025-AS011016445 on the supercomputer Jean Zay’s H100 partition. We gratefully acknowledge support from GENCI and IDRIS and from Stephane Requena (GENCI) and Pierre-François Lavallée (IDRIS) in particular.

Luciole-23B-Audio-1.0 was created by members of LINAGORA, including in alphabetical order:

Audran Bert (LINAGORA)

Hédi Naouara (LINAGORA)

Jérôme Louradour (LINAGORA)

We thank the support teams from IDRIS and NVIDIA for technical guidance throughout the project.

Contact

contact@openllm-france.fr