Low-WER TTS for Reading Product Codes and Numbers

In the OpenMOSS Seed-TTS-eval table, MOSS-TTS 8B reports 1.84% English WER. The listed Qwen3-TTS 1.7B and 0.6B checkpoints report 1.50% and 1.68% respectively. That table supports a shortlist, with MOSS-TTS v1.5 offering pronunciation and pause controls for scripts that need them. OpenMOSS comparison · v1.5 model card

Extend the public benchmark to identifiers

Seed-TTS-eval transcribes synthesized speech and scores text edits against the reference. For a parts-catalog pilot, select adjacent items whose codes differ by one digit or letter. Ask listeners to write the codes down. A changed digit can send a listener to the wrong part even when the rest of the sentence is accurate. Benchmark implementation

Approve acceptable spoken forms before testing. Use the same recognizer, references, and normalization across models. Where the recognizer misreads an audible code, retain the automated score and annotate the listening result. Preserve the original audio and the exact model revision so the result can be reproduced.

OpenMOSS documents the 8B evaluation separately from its current v1.5 controls. This distinction matters when a team reports a new pilot. Use the approved script and actual measured outcome in the release record. 8B checkpoint · Official model page

See WER evaluation basics and help center narration.

English SEO keywords

lowest WER TTS, exact identifier accuracy, product codes.