Meta Text-AB Reality Check: 3B Parameters, 480K Hours, 4.53/5 Human-Likeness—and No Public Checkpoint Yet

Meta published Alignment-Free Text-Audiobox (Text-AB) on September 6, 2026, three days after the associated arXiv submission. The system is a research speech-generation framework for cross-lingual voice dubbing, full-duplex two-speaker dialogue and emotion-controlled dialogue synthesis. It is not a general-purpose chat model, and Meta does not currently present it as a public hosted API or downloadable checkpoint.

The most important claim is architectural rather than a single leaderboard score. Text-AB uses a 3-billion-parameter Diffusion Transformer trained with flow matching, represents 48 kHz audio as 25 Hz DAC-VAE latents, and learns text-to-speech alignment through cross-attention rather than requiring a forced aligner or explicit duration predictor. The paper says the DAC-VAE representation compresses audio by about 1,920× versus roughly 160×–320× for the EnCodec representations used by earlier Audiobox work.

Training scale is large, but the data are not public

The main Text-AB model was pretrained on approximately 480,000 hours of monolingual speech: about 380,000 hours of English and 100,000 hours of Spanish. Meta then reports roughly 2.1k hours of dubbing fine-tuning data, including 50 hours of synthetic cross-lingual English↔Spanish data, and 28,000 hours of two-channel English dialogue data for the full-duplex system.

The training recipe is also unusually compute-heavy for a speech model. The paper reports 800,000 pretraining steps on 256 NVIDIA A100 GPUs, 200,000 dubbing fine-tuning steps on 256 A100s, and 1 million dialogue fine-tuning steps on 256 A100s. Those numbers describe the research training setup; Meta does not publish a dollar training cost, energy figure or reproducible public training corpus in the release.

Voice-dubbing benchmark: 200 internal samples and a private baseline

For cross-lingual dubbing, the evaluation uses an internal real-world benchmark containing 100 English→Spanish and 100 Spanish→English samples. Human raters compare Text-AB with Meta’s latest internal dubbing system on a -3 to +3 preference scale.

Text-AB is preferred on overall shareability by +0.40 for Spanish→English and +0.38 for English→Spanish. The paper also reports gains in prosody similarity (+0.33/+0.34), voice similarity (+0.29/+0.36), and voice naturalness (+0.39/+0.45).

These are meaningful first-party results, but they are not an independently reproducible public benchmark. The dataset and competing internal system are not released, so outside researchers cannot currently replay the exact comparison or verify whether the same preference gap survives on a different domain, language pair or rater pool.

The headline quality also benefits from best-of-many reranking

Text-AB is stochastic, and Meta evaluates it with multi-stage reranking. For voice dubbing, the default setup generates 32 candidates, filters them by speaker similarity and then selects the remaining candidate with the lowest Whisper-Large-V3 word error rate.

The paper’s own ablation shows why this matters: moving from no reranking to 32 candidates improves WER from 4.05% to 2.20% and speaker similarity from 0.66 to 0.74. That is a real quality improvement, but it also means the strongest reported inference setting is not equivalent to one cheap single-pass generation. More candidates require more model executions and therefore more compute. The paper does not publish a per-minute dollar cost for this setting.

Model scaling also helps sharply in the paper’s dubbing ablation. A 300M model records 44.45% WER, a 1B model 22.97%, and the 3B model 13.98%, while speaker similarity rises from 0.64 to 0.76. These figures come from the authors’ own evaluation configuration and should not be mixed with the reranked 2.20% WER result, which is a different ablation.

Full-duplex dialogue: near ground truth on short clips, but not on every dimension

For short-form dialogue, Text-AB receives 4.53/5 human-likeness versus 4.62/5 for ground-truth recordings, a difference of 0.09 points. The short-form set averages roughly 30 seconds and comes from a held-out subset of the same overall training-data source. The paper does not provide a confidence interval for the 4.53-versus-4.62 aggregate, so “near ground truth” is fairer than claiming statistical parity.

On longer 1–2 minute prompts, Text-AB reaches 3.86/5 human-likeness versus 3.00/5 for Meta’s latest internal dialogue system. It also scores better on intonation, pacing, expressive intensity and expressive correctness. However, the internal baseline actually scores higher on non-speech vocalizations and fillers: 4.57 versus 4.12. That exception is important because the system does not dominate every conversational dimension.

The public paper describes the short- and long-form test construction but does not state the number of examples in those two dialogue sets. That missing sample count limits how confidently readers should generalize the MOS differences.

Emotion control shows gains, but the evaluators matter

For emotional dialogue, Meta constructs 200 multi-turn dialogues with an internal text model and assigns turn-level angry, happy, neutral or sad labels. Emotion accuracy is then judged with Qwen2-Audio, while an additional dialogue-naturalness score is produced by Gemini 2.5 Pro. Human MOS ratings are also collected.

The emotion-conditioned system improves objective emotion classification over the non-conditioned dialogue model for angry, happy and sad turns, and it raises human emotion-alignment MOS substantially. But the comparison is not uniformly one-sided. On ground-truth text prompts, MoonCast receives a slightly higher human naturalness MOS (3.990) than emotion-conditioned Text-AB (3.906), while Text-AB has the higher LLM-judge score. On synthetic emotional text, Text-AB has the best human MOS among the shown systems (3.866).

That disagreement is useful: model-based judges and human listeners do not always rank speech systems identically, so neither score should be treated as a universal measure of conversational quality.

Long-form support is a generation technique, not an unlimited context window

Text-AB can generate roughly one minute of speech in a single pass before quality degrades. Longer output uses a multi-diffusion scheme with overlapping chunks; the default long-form setup uses 30-second chunks with 20-second overlap. Meta says this supports generation beyond 10 minutes and, in principle, arbitrarily long audio.

That is different from an LLM-style token context window. There is no published “1M token context” equivalent for Text-AB. Long-form capability comes from iterative overlapping generation, and the authors note that mono or empty prompts can suffer speaker-identity drift across chunks unless a stereo prompt is established.

The paper also says increasing ODE steps and reranking candidates generally improves WER and audio quality while increasing real-time factor. It does not publish a single production latency number that can be generalized to a public serving endpoint.

Availability, pricing and reproducibility

As verified on September 7, Meta’s publication page and the arXiv paper expose the research description and evaluation, but no public Text-AB model checkpoint, hosted inference API, rate card or official latency SLA was confirmed. Therefore there is no responsible way to quote a Text-AB API price or claim a downloadable production release.

The paper provides substantial implementation detail, including parameter count, training hours, GPU count, ODE steps, reranking sizes and evaluation metrics. Reproduction is still constrained by the private training corpora, private dubbing benchmark and private internal comparison systems.

SWE-bench Verified and SWE-bench Pro are not Text-AB benchmarks

Text-AB is a speech-synthesis system, not a repository-editing coding agent. The Meta release and paper provide no SWE-bench Verified result and no SWE-bench Pro result. Those benchmark families should not be inferred from the model’s audio results, and no score from another Meta model should be transferred to Text-AB.

What public feedback currently tells us

Fresh searches for direct Text-AB discussion on X and Reddit did not surface a reproducible user benchmark, controlled A/B test or attributable performance study that could be treated as independent evidence. Because the Meta publication is new, it is safer to report that public feedback is still sparse than to manufacture a consensus from unrelated Audiobox discussions.

Practical tradeoffs

Text-AB’s strongest ideas are technically interesting: removal of forced alignment, a highly compressed 48 kHz latent representation, joint stereo dialogue generation, controllable emotion, and long-form multi-diffusion. The main caveats are equally important. The strongest benchmarks are internal, some evaluation uses automated model judges, several headline gains depend on best-of-many reranking, and there is no confirmed public checkpoint or API with measured price and latency.

For researchers, the next evidence to watch is straightforward: a released checkpoint or authorized API, a public test set, independent replay of the dubbing and full-duplex results, single-sample versus reranked cost-per-quality measurements, and broader multilingual evaluation beyond English and Spanish.

Primary sources: Meta AI Research’s September 6 publication and arXiv:2609.03992, submitted September 3, 2026.