Speaker-1-Lite

Speaker-1-Lite is an experimental, single-speaker English text-to-speech (TTS) system. It combines a Tacotron 2 acoustic model with a HiFi-GAN V3-style neural vocoder to generate 22.05 kHz speech from text.

Both networks were trained from random initialization. No pretrained TTS model or pretrained vocoder weights were used. The training data reported for this project is LJSpeech 1.1 only.

Release status: This card is based on an audit collected on November 10, 2026 at 16:33 CET. At that time, the newest saved HiFi-GAN checkpoint was at 171,000 steps and the best-named checkpoint was at 154,000 steps. The Tacotron 2 checkpoints were at 37,000 steps. This is therefore a checkpoint snapshot, not proof that the planned training runs had reached their final step counts. Update the checkpoint status after the final training run and before treating this card as a final training report.

Audio sample

The following is the supplied end-to-end sample generated with the Tacotron 2 + HiFi-GAN path. The text prompt used for this particular WAV is: "Hello There. This is Speaker One Lite. I am a Text to Speech Model."

Download the sample WAV

Model details

Field	Value
Model name	Speaker-1-Lite
Task	Text-to-speech
Language	English
Speaker setup	Single speaker
Acoustic model	Tacotron 2
Neural vocoder	HiFi-GAN V3-style generator (ResBlock2)
Sample rate	22,050 Hz
Output audio	Mono WAV
Mel channels	80
FFT / window length	1,024 / 1,024
Mel hop length	256 samples
Pretrained weights	None; trained from random initialization

The inspected checkpoints contain approximately 7.24 million Tacotron 2 parameters and 1.46 million HiFi-GAN generator parameters. The generator count does not include the vocoder's training discriminators or optimizer state.

Architecture

Tacotron 2 acoustic model

• Character/token embedding dimension: 256
• Encoder: 3 convolutional layers, kernel size 5
• Decoder RNN dimension: 512
• Attention RNN dimension: 512
• Attention hidden dimension: 128
• Location-sensitive attention: 32 filters, kernel size 31
• Prenet dimension: 128
• Postnet: 5 convolutional layers, embedding dimension 256, kernel size 5
• Decoder maximum steps: 860
• Gate threshold: 0.5

HiFi-GAN vocoder

• HiFi-GAN V3-style generator using ResBlock2 residual blocks
• Initial upsampling channels: 256
• Upsampling rates: [8, 8, 4]
• Upsampling kernel sizes: [16, 16, 8]
• Residual block kernel sizes: [3, 5, 7]
• Residual block dilation pairs: [[1, 2], [2, 6], [3, 12]]
• Multi-period and multi-scale discriminators are used during training.

Training data

The reported training dataset is LJSpeech 1.1. Its official dataset page describes version 1.1 as public domain and lists 13,100 single-speaker audio clips with approximately 24 hours of speech.

The local project audit found:

Dataset inventory	Count / duration
Source WAV files present	13,100
Total source WAV duration	23.921 hours
Training manifest records	11,754
Validation manifest records	240
Total records in the two manifests	11,994
Precomputed mel files	11,994

The manifests, rather than the full directory file count, identify the examples prepared for the training and validation loaders. The audit does not establish why the remaining source WAV files are absent from those manifests.

The project configuration specifies a 2% validation fraction and filters utterances using a minimum duration of 1.0 second, maximum duration of 9.5 seconds, and maximum text length of 190 characters.

No additional datasets were reported by the model owner.

Training information

The following is the state observed in the audit report. It should not be interpreted as a complete training log.

Component	Checkpoint observed	Steps recorded	Notes
Tacotron 2	checkpoints/latest.pt and checkpoints/best.pt	37,000	Both checkpoint files reported the same step
HiFi-GAN	vocoder_checkpoints/latest.pt	171,000	Newest checkpoint found at audit time
HiFi-GAN	vocoder_checkpoints/best.pt	154,000	Best-named checkpoint found at audit time

Tacotron 2 configuration found in config.json:

• Batch size: 16
• Gradient accumulation: 1
• Optimizer learning rate: 0.0003
• Weight decay: 0.000001
• Precision: FP16
• Configured maximum training steps: 100,000
• Gradient clipping: 1.0
• Validation interval: 1,000 steps
• Checkpoint interval: 500 steps
• Random seed: 1337

HiFi-GAN settings confirmed by the project audit / owner report:

• Reported batch size: 2
• Training segment length: 8,192 waveform samples (approximately 0.372 seconds at 22,050 Hz)
• Vocoder implementation reports random initialization and no pretrained checkpoints.

Training logs were not present in the audited project folder. Therefore, this card does not report final loss values, loss curves, total training time, or best validation loss.

Hardware and software environment

The audit detected the following environment:

Component	Version / device
Operating system	Windows 10 (build 26200)
Python	3.11.9
PyTorch	2.5.1+cu121
TorchAudio	2.5.1+cu121
CUDA runtime	12.1
cuDNN	9.1.0
GPU	NVIDIA GeForce RTX 4060, 8 GB VRAM
NumPy	1.26.4
SoundFile	0.12.1

These are the versions found in the training environment, not a claim that all versions are required for inference.

Inference

The project includes vocoder_synthesize.py for end-to-end inference. From the project directory, with the source files, config.json, and the selected checkpoints in the expected locations:

python -m pip install -r requirements.txt

python vocoder_synthesize.py \
  --text "Hello there." \
  --tacotron-checkpoint checkpoints/best.pt \
  --vocoder-checkpoint vocoder_checkpoints/best.pt \
  --out outputs/hello_hifigan.wav

On Windows PowerShell, the same command can be entered on one line:

python vocoder_synthesize.py --text "Hello there." --tacotron-checkpoint checkpoints/best.pt --vocoder-checkpoint vocoder_checkpoints/best.pt --out outputs/hello_hifigan.wav

The inference script expects the project configuration and checkpoint format produced by this repository. The command above is for the repository's local project layout; it is not a standalone transformers or hosted Inference API example.

Evaluation

No formal human or automatic speech-quality evaluation was included in the audit. In particular, the project report contains no MOS, CMOS, WER, MCD, speaker-similarity score, or other validated quality metric. Training logs were also not available, so no loss curve is published here.

The project contains two different kinds of sample:

• vocoder_samples/ are HiFi-GAN reconstructions from real validation mel spectrograms. They assess the vocoder's waveform reconstruction path and are not end-to-end text-to-speech samples.
• samples/ are Tacotron 2 outputs from the fixed text prompts in the training configuration.
• outputs/ contains end-to-end synthesis outputs. The attached sample in this repository is one such Tacotron 2 + HiFi-GAN output.

One sample is not enough to establish pronunciation accuracy, intelligibility across prompts, speaker similarity, prosody, or generalization. No benchmark result is claimed.

Intended use

Speaker-1-Lite is intended for:

• TTS experimentation and educational use;
• prototyping and evaluation of a small, single-speaker neural TTS pipeline;
• research into acoustic-model and neural-vocoder training workflows.

It is not presented as a production-grade TTS service.

Limitations and responsible use

• The model is designed for English and one speaker. It is not a multilingual or multi-speaker system.
• Output quality may vary with sentence length, punctuation, uncommon words, numbers, abbreviations, and text outside the training distribution.
• Pronunciation, alignment, duration, and prosody have not been systematically benchmarked.
• Synthetic speech may contain audible artifacts. Users should listen to generated output before relying on it.
• The supplied audio example is only one test sample; it should not be treated as a representative quality guarantee.
• Do not use generated speech to impersonate a real person, mislead listeners, or create deceptive recordings. Disclose synthetic generation where appropriate.

License

The model is released by its owner under the Apache License 2.0. The repository should include the complete Apache-2.0 license text in a LICENSE or LICENSE.txt file.

The LJSpeech 1.1 dataset page identifies the dataset as public domain. The dataset's status is separate from the license of this model and its implementation. Dependency licenses also remain separate; distributing this model under Apache-2.0 does not relicense third-party dependencies.

Dataset reference

Keith Ito. The LJ Speech Dataset, 2017. https://keithito.com/LJ-Speech-Dataset/