VALL-E X
No estimate
No hardware requirements for this model
The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.
On record
Full specification
Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.
Origin
Who built this model, where, and when it was published.
- Organisation
- Microsoft
- Organisation type
- Industry
- Country
- United States of America
- Published
- 7 March 2023
- Authors
- Ziqiang Zhang, Long Zhou, Chengyi Wang, Sanyuan Chen, Yu Wu, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, Lei He, Sheng Zhao, Furu Wei
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Speech
- Task
- Translation, Speech synthesis, Speech recognition (ASR), Speech-to-speech
Size
How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.
- Parameters
- 700M
- Training data
- 151,200,000,000 tokens
"For the cross-lingual codec language models, φMAR and φMNAR are both 12-layer Transformer decoders with an attention dimension of 1024 and the FFN dimension of 4096." These are two parts of the model. According to Ben's script, that's 353M parameters, or ~700M for both (https://github.com/bencottier/ml-parameter-count/blob/main/parameter_count.py)
70k hours, encoded by an 8 layer EnCodec, where each layer generates 75 encodings per second of 24 kHz audio. 70k hr * 3600 sec/hr * 75/sec = 18.9B audio encodings per layer φMAR is trained to autoregressively predict the each subsequent token in the first layer of encodings. φMNAR is trained to predict the remaining acoustic embedding layers after being shown both the first layer of the current input and the full set of embeddings from a previous sentence. Based on this, the total number of …
Training compute
The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.
- Training compute
- 1.2 × 10²¹ FLOP
- How it was established
- Operation counting
" The batch sizes of speech and phonemes on each GPU are 1,400,000 (87.5 seconds) and 3,000, respectively. The maximum learning rate is 5e-4 with warm-up steps of 32,000. The model is pre-trained on 32 V100 GPUs for 400K step" 6CN: 6 * (1.4M + 3000) * 353M * 400k = 1.2e21 (2 OOM larger than our estimate for VALL-E, and would take ~46 days on the V100 cluster at 30% utilization. Seems perhaps a bit high?)
The training run
What it physically took to train: which chips, how many, for how long, and what that drew from the wall.
- Training hardware
- NVIDIA V100
Availability
Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.
- Weights
- Closed — provider access only
- Model access
- Unreleased
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Record confidence
- Likely
- Citations
- 262
Sources
Where this record came from and when it was last checked.
- Reference
- Speak Foreign Languages with Your Own Voice: Cross-Lingual Neural Codec Language Modeling
- Last updated
- 25 May 2026
What the numbers mean
About this model
VALL-E X was published by Microsoft, in United States of America, in March 2023. The organisation is categorised as industry.
It works in Speech, and is recorded as doing translation, Speech synthesis, Speech recognition (ASR), Speech-to-speech.
This is a closed model: the trained values stayed with whoever produced them, and there is no local version to run.
What went into building it
Training it took roughly 1.2 × 10²¹ FLOP of computation, on NVIDIA V100 — a measure of what producing the model cost, not of how fast it answers.
It was trained on about 151,200,000,000 tokens of text.
Answers
VALL-E X — common questions
What GPU do I need to run VALL-E X?
None. VALL-E X is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.
Is VALL-E X open source?
No. VALL-E X has not had its weights published, so it exists only as a service controlled by its owner.
How many parameters does VALL-E X have?
VALL-E X has 700M parameters. "For the cross-lingual codec language models, φMAR and φMNAR are both 12-layer Transformer decoders with an attention dimension of 1024 and the FFN dimension of 4096." These are two parts of the model. According to Ben's script, that's 353M parameters, or ~700M for both (https://github.com/bencottier/ml-parameter-count/blob/main/parameter_count.py). That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.
Who created VALL-E X?
VALL-E X was published by Microsoft, based in United States of America, categorised as industry.
When was VALL-E X released?
VALL-E X was published in March 2023. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.
What is VALL-E X used for?
VALL-E X works in Speech, and is recorded as handling translation, Speech synthesis, Speech recognition (ASR), Speech-to-speech. These are the areas it was designed around; they describe intent rather than a hard boundary.
How much compute was used to train VALL-E X?
Around 1.2 × 10²¹ FLOP, on NVIDIA V100. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.
The other direction
Looking at it from the other side?
This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.