OptiSpeech-ConvNext-TTS-es-CO-Marcia-v1

OptiSpeech ConvNext-TTS es-CO-Marcia v1

OptiSpeech ConvNext-TTS es-CO-Marcia v1 is a Spanish single-speaker text-to-speech ONNX export trained with the OptiSpeech framework. The speaker is Marcia, targeting Colombian Spanish (es-CO).

This model uses the OptiSpeech Spanish TextProcessor at inference time. No separate gruut_*_symbols.py file is needed for this export.

Model

Files

• convnext-tts-es-CO-Marcia-v1.onnx: ONNX text-to-waveform export.
• config_tree.log: rendered training configuration.
• training_metadata/config.yaml: Hydra training config.
• training_metadata/hydra.yaml: Hydra runtime config.
• training_metadata/overrides.yaml: command-line overrides for the source run.
• onnx_inference_es.py: Spanish ONNX inference helper.

Training Source

This export was staged from:

logs/train/es_speaker_ddp_v2/convnext-tts-ES-Marcia.onnx
logs/train/es_speaker_ddp_v2/runs/2026-06-30_13-54-31/

The source run used:

experiment=es_speaker_ddp_finetune
run_name=es_speaker_ddp_v2

ONNX Inference

Run inference from an OptiSpeech checkout with the Spanish text processor dependencies installed:

source .venv/bin/activate
uv sync --extra es

python onnx_inference_es.py \
  --model_path convnext-tts-es-CO-Marcia-v1.onnx \
  --input "Hola, esta es una prueba de síntesis de voz en español colombiano." \
  --output_path outputs \
  --output_name marcia_v1.wav

The ONNX model expects:

• x: phoneme/token IDs from optispeech.text.TextProcessor(languages=["es"])
• x_lengths: input sequence lengths
• scales: duration, pitch, and energy factors

The helper script handles text normalization, Spanish phonemization/tokenization, ONNXRuntime execution, and WAV writing.

Framework

• OptiSpeech
• ONNXRuntime
• PyTorch / Lightning training stack