When inspecting the quality of a speech synthesis model using only STT, we missed all added sounds lasting just 0.1 seconds after silence. I’ll explain how we built a detector using waveform envelopes and how all 12 models produced false positives due to commas.
Quality Control for TTS Models: Why Transcription Isn’t Enough
I used to perform quality control (QC) for TTS models using this process:
Have the model read probe sentences
Transcribe with Whisper
Compare against the script to check accuracy and trailing elongation
Word Error Rate lies in both directions on text-to-speech. A 390-sample blind study: 55% of the uncertain clips were the…
The trimming process designed to remove hallucinated clips from the speech synthesis training corpus was 'inadvertently disabled…
A collection of real-world techniques that improved faster-whisper performance in live voice interactions. Covers proper use of…
In the selection of voice model candidates, we included 'resistance to generation breakdown' as an evaluation metric, but that…
Voice AI agents are magical until they start lagging. Suddenly my Groq-powered voice assistant...
I tried to implement a diffusion-based TTS that generates voices from captions for conversational use, but it was 2.5 times…
