Llasa.
1B · 3B · 8B.
Three official Llasa base models: 1B, 3B and 8B. Dia2 1B/2B and the Dia 1.6B reference control also remain in this evaluation. Full evaluations use all 1,088 English Seed-TTS-Eval texts.
Original Dia and Llasa runs had no reference audio. This experiment holds an English reference voice fixed and tests whether conditioning changes intelligibility.
Every checkpoint, accounted for.
Loading verified progress…
| Checkpoint | Pilot · 8 texts | Full · 1,088 texts | Full WER | Full CER | Evidence |
|---|
Completed full runs
Only complete, verified 1,088-text results enter this chart. Click a model for recordings. ‡ marks no-audio failures included in corpus scores.
Does reference audio change the result?
Paired controls use the same eight target texts, evaluator, seed and model implementation. Reference conditioning is the changed factor.
Listen to each result
Targets, Whisper transcripts, hash-verified audio and explicit generation failures. Pilot recordings remain clearly marked.
Frozen protocol
English only · seed 42 · one generation per target. Corpus WER/CER aggregate edit counts. Whisper-large-v3 (Systran conversion, revision edaa852ec7e145841d8ffdb056a99866b5f0a478), CUDA FP16, beam 5, temperature 0, no VAD, no previous-text conditioning; whisper_english normalization. Full-run confidence intervals use 1,000 prompt bootstrap resamples.
Llasa: independent Transformers LM/tokenizer, BF16, temperature 0.8, top-p 1, top-k 50; frozen independently audited XCodec2 codec in FP32. Dia: independent Transformers processor, LM and DAC in FP32. Dia2: official Nari Labs implementation, BF16, CFG 2, audio temperature 0.8 and top-k 50; fixed prefix word alignment is prepared once with Whisper-large-v3. Reference audio is excluded from scored output.
The active Llasa selection is HKUSTAudio/Llasa-1B, HKUSTAudio/Llasa-3B and HKUSTAudio/Llasa-8B. Multilingual, Preserve-TextChat and multi-speaker variants have been removed from the active evaluation. Dia and both Dia2 checkpoints remain selected; every checkpoint is pinned to a commit. Community fine-tunes and duplicate quantizations are outside this inventory. Fixed reference speech differs from the official per-sample speaker-similarity protocol.
Official Llasa inventory ↗ · Official Dia2 code ↗ · Versioned experiment artifacts ↗