speecht5_tts-finetuned-nst-da

This model is a fine-tuned version of microsoft/speecht5_tts on the NST Danish ASR Database dataset. It achieves the following results on the evaluation set:

• Loss: 0.3692

Model description

Given that danish is a low-resource language, not many open-source implementations of a danish text-to-speech synthesizer are available online. As of writing, the only other existing implementations available on 🤗 are facebook/seamless-streaming and audo/seamless-m4t-v2-large. This model has been developed to provide a simpler alternative that still performs reasonable well, both in terms of output quality and inference time. Additionally, contrary to the aforementioned models, this model also has an associated Space on 🤗 at JackismyShephard/danish-speech-synthesis which provides an easy interface for danish text-to-speech synthesis, as well as optional speech enhancement.

Intended uses & limitations

The model is intended for danish text-to-speech synthesis.

The model does not recognize special symbols such as "æ", "ø" and "å", as it uses the default tokenizer of microsoft/speecht5_tts. The model performs best for short to medium length input text and expects input text to contain no more than 600 vocabulary tokens. Additionally, for best performance the model should be given a danish speaker embedding, ideally generated from an audio clip from the training split of alexandrainst/nst-da using speechbrain/spkrec-xvect-voxceleb.

The output of the model is a log-mel spectogram, which should be converted to a waveform using microsoft/speecht5_hifigan. For better quality output the resulting waveform can be enhanced using ResembleAI/resemble-enhance.

An example script showing how to use the model for inference can be found here.

Training and evaluation data

The model was trained and evaluated on alexandrainst/nst-da using MSE as both loss and metric. The dataset was pre-processed as follows:

• special characters, such as "æ", "ø" and "å" were translated to their latin equivalents and examples with text containing digits were removed, as neiher are in the vocabulary of the tokenizer of microsoft/speecht5_tts.
• training split balancing was done by excluding speakers with less than 280 examples or more than 327 examples.
• audio was enhanced using speechbrain/metricgan-plus-voicebank in an attempt to remove unwanted noise.

Training procedure

The script used for training the model (and pre-processing its data) can be found here.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
0.4445	1.0	9429	0.4100
0.4169	2.0	18858	0.3955
0.412	3.0	28287	0.3882
0.3982	4.0	37716	0.3826
0.4032	5.0	47145	0.3817
0.3951	6.0	56574	0.3782
0.3971	7.0	66003	0.3782
0.395	8.0	75432	0.3757
0.3952	9.0	84861	0.3749
0.3835	10.0	94290	0.3740
0.3863	11.0	103719	0.3754
0.3845	12.0	113148	0.3732
0.3788	13.0	122577	0.3715
0.3834	14.0	132006	0.3717
0.3894	15.0	141435	0.3718
0.3845	16.0	150864	0.3714
0.3823	17.0	160293	0.3692
0.3858	18.0	169722	0.3703
0.3919	19.0	179151	0.3716
0.3906	20.0	188580	0.3709

Framework versions

• Transformers 4.37.2
• Pytorch 2.1.1+cu121
• Datasets 2.17.0
• Tokenizers 0.15.2