speecht5-tts-ov

Microsoft SpeechT5 text-to-speech with the SpeechT5 HiFi-GAN vocoder, exported to OpenVINO IR (fp32, no quantisation) for use with RustedVINO and OpenVINO GenAI's Text2SpeechPipeline.

• Language: English. Output: 16 kHz mono. The voice is plain and somewhat robotic; this is the smallest English voice in RustedVINO's starter set, not a high-quality one.
• Licence: MIT, inherited from both upstream models. All credit for the model goes to Microsoft (Ao et al., SpeechT5: Unified-Modal Encoder-Decoder Pre-Training for Spoken Language Processing, 2022). This repository only contains a format conversion.

How it was made

optimum-cli export openvino --model microsoft/speecht5_tts \
  --model-kwargs '{"vocoder": "microsoft/speecht5_hifigan"}' --weight-format fp32 speecht5-tts-ov
# tokenizer files saved from microsoft/speecht5_tts, then:
convert_tokenizer speecht5-tts-ov --with-detokenizer -o speecht5-tts-ov

Python 3.12, optimum-intel 2.0.0, optimum 2.2.0, transformers 4.57.6, OpenVINO 2026.2.1, openvino-tokenizers 2026.2.1.0.

Use with RustedVINO

hf download jermalk/speecht5-tts-ov --local-dir /opt/rustedvino/models/speecht5-tts-ov
rv add speecht5-tts-ov --kind tts
curl -s http://localhost:11437/v1/audio/speech -H "Content-Type: application/json" \
  -d '{"model":"speecht5-tts-ov","input":"Hello from RustedVINO.","language":"en"}' -o hello.wav

Measured on one Intel Core Ultra 7 laptop, CPU only: about 0.76× real time (a 5-second sentence in about 4 seconds), roughly 2.5 GB of RAM in use while synthesising.