Model reference · open weights

xls-r-sv-cv8

Audio masapasa · community Speech→text 1 build Open weights 722 dl/mo

xls-r-sv-cv8 is an open-weight audio or speech model from masapasa. xls-r-300m-sv-cv8 (BF16) weighs 1.3 GB; the smallest configuration that runs it is RTX 3060 12 GB.

Hugging Face ↗

What it is

Released by	masapasa
Type	Audio & music
Task	Speech→text
Runs with	transformers
Released	2022-03-02
Popularity	722 downloads / month
Weights	1.3 GB (xls-r-300m-sv-cv8 (BF16), file size)
Licence	Open weights

What it runs on

Memory and cards for xls-r-300m-sv-cv8 (BF16)

Weights 1.3 GB (file size) · overhead about 1.6 GB.

Card	One stream	Counted memory
RTX 3060 12 GB	fits	11.6 GB
RTX 4060 Ti 16 GB	fits	15.4 GB
RTX 3090 24 GB	fits	23.4 GB
RTX 4090 24 GB	fits	23.4 GB
RTX 5090 32 GB	fits	31.0 GB
L40S 48 GB	fits	44.0 GB
A100 80 GB	fits	78.2 GB
H100 80 GB	fits	78.1 GB
RTX PRO 6000 Blackwell 96 GB	fits	93.8 GB
DGX Spark (GB10) 128 GB unified	fits	107 GB
H200 141 GB	fits	138 GB
B200 180 GB	fits	176 GB

Estimates, not measurements: the weights are the build's file size. A speech model's decoder keeps a small cache for every stream it transcribes, so memory grows with the streams and beams at once. Counted memory is 92 % of what CUDA reports for the card.

From the model card

What masapasa says about xls-r-sv-cv8

It achieves the following results on the evaluation set:

Loss: 2.3347
Wer: 1.0286

Model description

More information needed

Intended uses & limitations

More information needed

Read the full model card

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 7.5e-05
train_batch_size: 8
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 20
num_epochs: 5.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
10.7838	0.01	5	14.5035	1.0
13.0582	0.03	10	13.6658	1.0
7.3034	0.04	15	9.7898	1.0
6.1847	0.05	20	6.9148	1.0
5.3371	0.07	25	5.3661	1.0
4.4274	0.08	30	4.6945	1.0
4.0918	0.1	35	4.3172	1.0
4.1734	0.11	40	4.0759	1.0
3.7332	0.12	45	3.9039	1.0
3.6871	0.14	50	3.7777	1.0
3.4428	0.15	55	3.6718	1.0
3.5514	0.16	60	3.5947	1.0
3.4307	0.18	65	3.5144	1.0
3.4102	0.19	70	3.4432	1.0
3.4964	0.21	75	3.3890	1.0
3.3936	0.22	80	3.3467	1.0
3.3051	0.23	85	3.3102	1.0
3.278	0.25	90	3.2801	1.0
3.2223	0.26	95	3.2440	1.0
3.1888	0.27	100	3.2900	1.0
3.218	0.29	105	3.2627	1.0
3.1308	0.3	110	3.2152	1.0
3.109	0.31	115	3.1686	1.0
3.1188	0.33	120	3.1734	1.0
3.1132	0.34	125	3.1431	1.0
3.0667	0.36	130	3.1686	1.0
3.1167	0.37	135	3.1885	1.0
3.0592	0.38	140	3.1100	1.0
3.0531	0.4	145	3.1149	1.0
3.1224	0.41	150	3.1205	1.0
3.0651	0.42	155	3.1101	1.0
3.0077	0.44	160	3.0980	1.0
3.0027	0.45	165	3.1132	1.0
3.0423	0.47	170	3.0886	1.0
3.0462	0.48	175	3.0865	1.0
3.0701	0.49	180	3.0863	1.0
3.0871	0.51	185	3.0825	1.0
3.0585	0.52	190	3.0720	1.0
3.0274	0.53	195	3.0736	1.0
3.0983	0.55	200	3.0658	1.0
3.0538	0.56	205	3.1241	1.0
3.0862	0.57	210	3.0573	1.0
3.0041	0.59	215	3.0608	1.0
3.027	0.6	220	3.0614	1.0
2.9916	0.62	225	3.0527	1.0
3.0157	0.63	230	3.0514	1.0
3.0429	0.64	235	3.0391	1.0
2.999	0.66	240	3.0462	1.0
3.0053	0.67	245	3.0438	1.0
2.9812	0.68	250	3.0447	1.0
3.0062	0.7	255	3.0660	1.0
3.0045	0.71	260	3.0103	1.0
2.9684	0.73	265	3.0106	1.0
2.9885	0.74	270	3.0014	1.0
3.0062	0.75	275	2.9885	1.0
2.9736	0.77	280	3.0330	1.0
2.9766	0.78	285	2.9910	1.0
2.9545	0.79	290	2.9972	1.0
2.9936	0.81	295	2.9872	1.0
3.0832	0.82	300	2.9978	1.0
2.974	0.83	305	2.9978	1.0
2.9846	0.85	310	2.9849	1.0
2.9554	0.86	315	2.9810	1.0
2.9524	0.88	320	2.9731	1.0
2.9426	0.89	325	2.9824	1.0
2.9416	0.9	330	2.9731	1.0
2.9705	0.92	335	2.9830	1.0
2.9502	0.93	340	2.9713	1.0
2.9393	0.94	345	2.9790	1.0
2.9336	0.96	350	2.9684	1.0
2.9542	0.97	355	2.9689	1.0
2.9408	0.98	360	2.9556	1.0
2.9544	1.0	365	2.9563	1.0
2.9187	1.01	370	2.9624	1.0
2.9935	1.03	375	2.9500	1.0
2.9803	1.04	380	2.9558	1.0
2.9867	1.05	385	2.9473	1.0
2.8925	1.07	390	2.9444	1.0
2.9633	1.08	395	2.9490	1.0
2.9191	1.1	400	2.9362	1.0
2.9081	1.11	405	2.9394	1.0
2.9381	1.12	410	2.9846	1.0
2.9271	1.14	415	2.9638	1.0
2.959	1.15	420	2.9835	1.0
2.9486	1.16	425	2.9361	1.0
2.9246	1.18	430	2.9615	1.0
2.923	1.1

Quoted from the model card on Hugging Face — the full card is behind the Hugging Face link above.