I recorded 1,180 TTS clips to find out what voices misread in LLM output

Description and use

Preprocessing LLM output with speechready reduces voice misread rates from 32% to 8% overall, and similarly in English and Turkish, without regressing control prose.

Problem

Voice agents often misread text generated by LLMs, especially numbers, dates, phone numbers, and code, leading to incorrect spoken output.

Outcome

Preprocessing with speechready reduced overall misread rate from 32% to 8%, with similar improvements in English and Turkish, and no regression in control prose.

Workflow

• 1. Collect LLM-generated text containing markdown, numbers, dates, phone numbers, URLs, and code.
• 2. Run speechready to clean, verbalize, compile, and chunk the text.
• 3. Synthesize the original and processed text with each TTS voice.
• 4. Transcribe both recordings with gpt-4o-transcribe.
• 5. Score the transcriptions against the intended words, forgiving transcription conventions.
• 6. Calculate misread rates and compare before/after.
• 7. Iterate on rules and provider profiles based on held‑out tests.

Evidence

I recorded 1,180 TTS clips to find out what voices misread in LLM output

Source

Result, on the same 1,180 recordings: All 32% → 8%

Source

The scorer forgives the ways a transcription engine writes speech down: digits for spoken numbers, "21st" for "twenty-first", "NPM" for "en pi em".

Source

Providers disagree on what they handle. Google's Chirp voices read 25°C fine; others do not.

Source