LinTO STT French Punctuated – FastConformer

Overview

This model is a fine-tuned version of the NVIDIA French FastConformer Hybrid Large model. It is a large (115M parameters) hybrid ASR model trained with both Transducer (default) and CTC losses.

Compared to the base model, this version:

• Was trained on 10,000+ hours of diverse, manually transcribed French speech with punctuation.

Performance

The evaluation code is available in the ASR Benchmark repository.

Word Error Rate (WER)

WER was computed without punctuation or uppercase letters and datasets were cleaned. The SUMM-RE dataset is the only one used exclusively for evaluation, meaning neither model saw it during training.

Evaluations can be very long (especially for whisper) so we selected only segments with a duration over 1 second and used a subset of the test split for most datasets:

• 15% of CommonVoice: 2424 rows (3.9h)
• 33% of MultiLingual LibriSpeech: 800 rows (3.3h)
• 33% of SUMM-RE: 1004 rows (2h). We selected only segments above 4 seconds to ensure quality.
• 33% of VoxPopuli: 678 rows (1.6h)
• Multilingual TEDx: 972 rows (1.5h)
• 50% of our internal Youtube corpus: 956 rows (1h)

As shown in the table above (lower is better), the model demonstrates robust performance across all datasets, consistently achieving results close to the best even on the out-of-domain dataset like SUMM-RE.

Usage

Usage with Transformers

# Install transformers (>= 5.0 required for ParakeetForRNNT)
pip install "transformers>=5" librosa soundfile

from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="linagora/linto_stt_fr_fastconformer_pc")

result = asr("/path/to/your/audio/file.wav")
print(result["text"])

Or, for batched inference :

import librosa
from transformers import AutoProcessor, ParakeetForRNNT

model_name = "linagora/linto_stt_fr_fastconformer_pc"
processor = AutoProcessor.from_pretrained(model_name)
model = ParakeetForRNNT.from_pretrained(model_name)

audios = [librosa.load(path, sr=16000)[0] for path in ["audio1.wav", "audio2.wav"]]
inputs = processor(audios, sampling_rate=16000, return_tensors="pt", padding=True)
outputs = model.generate(**inputs)
print(processor.batch_decode(outputs.sequences, skip_special_tokens=True))

Usage with NeMo

This model can be used with the NVIDIA NeMo Toolkit for both inference and fine-tuning.

pip install nemo_toolkit['all']

import nemo.collections.asr as nemo_asr

model_name = "linagora/linto_stt_fr_fastconformer_pc"
asr_model = nemo_asr.models.ASRModel.from_pretrained(model_name=model_name)

# Path to your 16kHz mono-channel audio file
audio_path = "/path/to/your/audio/file"

# Transcribe with defaut transducer decoder
asr_model.transcribe([audio_path])

# (Optional) Switch to CTC decoder
asr_model.change_decoding_strategy(decoder_type="ctc")

# (Optional) Transcribe with CTC decoder
asr_model.transcribe([audio_path])

Usage in LinTO STT

It can also be used with the LinTO STT API, an Automatic Speech Recognition (ASR) API that can function either as a standalone transcription service or be deployed within a microservices infrastructure using a message broker connector. It supports both offline and real-time transcription.

Training Details

The training code is available in the nemo_asr_training repository.
The full configuration used for fine-tuning is available here.

Hardware

• 1× NVIDIA H100 GPU (80 GB)

Training Configuration

• Precision: BF16 mixed precision
• Max training steps: 100,000
• Gradient accumulation: 4 batches

Tokenizer

• Type: SentencePiece
• Vocabulary size: 1,024 tokens

Optimization

• Optimizer: AdamW
  • Learning rate: 1e-5
  • Betas: [0.9, 0.98]
  • Weight decay: 1e-3
• Scheduler: CosineAnnealing
  • Warmup steps: 10,000
  • Minimum learning rate: 1e-6

Data Setup

• Audios ranging from 0.1s to 30s
• Batch size of 50
• All datasets except YODAS and TouTubeFR were upsampled 2x

Training datasets

The data were transformed, processed and converted using NeMo tools from the SSAK repository. YODAS segments were merged to make segments lasting up to 30. This was done in order to improve the model capabilities on longer segments.

The model was trained on over 10,000 hours of French speech, covering:

• Read and spontaneous speech
• Conversations and meetings
• Varied accents and audio conditions

Datasets Used (by size):

• YouTubeFr: an internally curated corpus of CC0-licensed French-language videos sourced from YouTube. Will soon be available on LeVoiceLab platform
• YODAS: fr000 subset
• Multilingual LibriSpeech: french subset
• CommonVoice: french subset
• ESLO
• VoxPopuli: french subset
• Multilingual TEDx: french subset
• TCOF
• CTF-AR (Corpus de Conversations Téléphoniques en Français avec Accents Régionaux): will soon be available on LeVoiceLab platform
• PFC
• OFROM
• CTFNN1 (Corpus de Conversations Téléphoniques en Français impliquant des accents Non-Natifs): will soon be available on LeVoiceLab platform
• CFPP2000
• VOXFORGE
• CLAPI
• AfricanAccentedFrench
• FLEURS: french subset
• LesVocaux
• LINAGORA_Meetings
• CFPB
• ACSYNT
• PxSLU
• SimSamu

Limitations

• May struggle with rare vocabulary, heavy accents, or overlapping/multi-speaker audio.
• May struggle with punctuations on long segments.

References

[1] Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition

[2] Google Sentencepiece Tokenizer

[3] NVIDIA NeMo Toolkit

Acknowledgements

Training of linto_stt_fr_fastconformer_pc was made possible by computing AI and storage resources by GENCI at IDRIS thanks to the grant 2025-A0181016189 on the supercomputer Jean Zay’s H100 partition.

Thanks to NVIDIA for providing the base model architecture and the NeMo framework.

Licence

The model is released under a CC-BY-4.0 license, in line with the licensing of the original model it was fine-tuned from.