Soloni TDT-CTC 114M Series

| |

soloni-114m-tdt-ctc-v3 is a fine tuned version of RobotsMali/soloni-114m-tdt-ctc-v2. The model was fine-tuned using NVIDIA NeMo and supports both TDT (Token-and-Duration Transducer) and CTC (Connectionist Temporal Classification) decoding.

🚨 Important Note

This model, along with its associated resources, is part of an ongoing research effort, improvements and refinements are expected in future versions. Users should be aware that:

• The model may not generalize very well across all speaking conditions and dialects.
• Community feedback is welcome, and contributions are encouraged to refine the model further.

NVIDIA NeMo: Training

To fine-tune or play with the model you will need to install NVIDIA NeMo. We recommend you install it after you've installed latest PyTorch version.

pip install nemo-toolkit['asr']

How to Use This Model

Note that this model has been released for research purposes primarily.

Load Model with NeMo

import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.EncDecHybridRNNTCTCBPEModel.from_pretrained(model_name="RobotsMali/soloni-114m-tdt-ctc-v3")

NeMo 2.5.0 Compatibility

This checkpoint was created with NeMo 2.5.0. Loading it with newer NeMo versions (observed with 2.7.x) can fail because the strict decoding schema expects key_phrase_items_list; see NVIDIA-NeMo/Speech#15658. This workaround was tested with Python 3.12:

from pathlib import Path

from nemo.collections.asr.models import ASRModel
from omegaconf import OmegaConf

model_name = "RobotsMali/soloni-114m-tdt-ctc-v3"
cfg = ASRModel.from_pretrained(model_name, return_config=True)
OmegaConf.set_struct(cfg, False)

for decoder in ("greedy", "beam"):
    boosting_tree = OmegaConf.select(cfg, f"decoding.{decoder}.boosting_tree")
    if boosting_tree is not None:
        boosting_tree.key_phrase_items_list = None

config_path = Path("patched_config.yaml").resolve()
OmegaConf.save(cfg, config_path)
asr_model = ASRModel.from_pretrained(
    model_name=model_name,
    override_config_path=str(config_path),
    strict=False,
)
config_path.unlink()

Transcribe Audio

# Assuming you have a test audio file named sample_audio.wav
asr_model.transcribe(['sample_audio.wav'])

Note that the decoding strategy for the TDT decoder use CUDA Graphs by default but not all GPUs and versions of cuda support that parameter. If you run into a RuntimeError: CUDA error: invalid argument you should set that argument to false in the decoding strategy before calling asr_model.transcribe()

decoding_cfg = asr_model.cfg.decoding
# Disable CUDA Graphs
decoding_cfg.greedy.use_cuda_graph_decoder = False
# Then change the decoding strategy
asr_model.change_decoding_strategy(decoding_cfg=decoding_cfg)

Input

This model accepts 16 kHz mono-channel audio (wav files) as input. But it is equipped with its own preprocessor doing the resampling so you may input audios at higher sampling rates.

Output

This model provides transcribed speech as an hypothesis object with a text attribute containing the transcription string for a given speech sample.

Model Architecture

This model uses a Hybrid FastConformer-TDT-CTC architecture. FastConformer is an optimized version of the Conformer model with 8x depthwise-separable convolutional downsampling. It possesses two independent but jointly trained decoders, one auto-regressive TDT decoder and a convolutional decoder with CTC loss. You may find more information on the details of FastConformer here: Fast-Conformer Model.

Training

The NeMo toolkit was used for finetuning this model for 13,388 steps over RobotsMali/soloni-114m-tdt-ctc-v2. The finetuning codes and configurations can be found at RobotsMali-AI/bambara-asr.

The tokenizer for this model was trained on the text transcripts of the train set of RobotsMali/kunkado using this script.

Dataset

This model was fine-tuned on the kunkado dataset, the human-reviewed subset, which consists of ~40 hours of transcribed Bambara speech data. The text was normalized with the bambara-normalizer prior to training, normalizing numbers, removing punctuations and removing tags.

Performance

The performance of Automatic Speech Recognition models is commonly measured using Word Error Rate (WER) and Character Error Rate (CER). Since this model has two decoders operating independently at inference time, each decoder is evaluated independently too.

The following table shows these two metrics for each decoder:

These are greedy WER numbers without external LM. By default the main decoder branch is the TDT branch, if you would like to switch to the CTC decoder simply run this block of code before calling the .transcribe method

# Retrieve the CTC decoding config
ctc_decoding_cfg = asr_model.cfg.aux_ctc.decoding
# Then change the decoding strategy
asr_model.change_decoding_strategy(decoder_type='ctc', decoding_cfg=ctc_decoding_cfg)
# Transcribe with the CTC decoder
asr_model.transcribe(['sample_audio.wav'])

These are greedy WER numbers without external LM.

License

This model is released under the CC-BY-4.0 license. By using this model, you agree to the terms of the license.

Feel free to open a discussion on Hugging Face or file an issue on GitHub for help or contributions.