Update README.md

README.md

@@ -1,9 +1,107 @@
-Model : /checkpoint/yilinyang/s2st/hok2en_fbank_mined/v1.mined_th1.07.rdrop_0.ls_0.1.maxtok_6.0k.uf_10.lr_0.0005.wu_2k.seed_1234.arch_xm_transformer_t2.frz_0.dr_0.1.ld_0.1.mcl_10.al_1.dld0.1.ca.mtl_config_t2_tug65k8.0.tdec_12L.synenc_2L.syndec_2L.ngpu16/checkpoint_last.pt 
+---
+license: cc-by-nc-4.0
+library_name: fairseq
+task: audio-to-audio
+tags:
+- fairseq
+- audio
+- audio-to-audio
+- speech-to-speech-translation
 
-Eval : python ust_common/scripts/generation/st/run_asr_bleu_pipeline.py data=/large_experiments/ust/yilinyang/data/s2st/hok2en gen_subset=dev_TAT_1hr config_yaml=config_t2_fbank.yaml unit_generation_args.max_tokens=6000 +unit_generation_args.multitask_config_yaml=mtl_config_t2_tug65k8.0_rdrop10.0.yaml +unit_generation_args.beam_mt=10 unit_generation_args.beam=10 unit_generation_args.max_len_a=0.3 +unit_generation_args.max_len_b=200 +unit_generation_args.max_len_a_mt=0 +unit_generation_args.max_len_b_mt=200 checkpoint_filename=checkpoint_last.pt experiment_dirpath=/checkpoint/yilinyang/s2st/hok2en_fbank_mined/v1.mined_th1.07.rdrop_0.ls_0.1.maxtok_6.0k.uf_10.lr_0.0005.wu_2k.seed_1234.arch_xm_transformer_t2.frz_0.dr_0.1.ld_0.1.mcl_10.al_1.dld0.1.ca.mtl_config_t2_tug65k8.0.tdec_12L.synenc_2L.syndec_2L.ngpu16
+datasets:
+- mtedx
+- covost2
+- europarl_st
+- voxpopuli
+widget:
+- example_title: Common Voice sample 1
+  src: https://huggingface.co/facebook/xm_transformer_600m-es_en-multi_domain/resolve/main/common_voice_es_19966634.flac
+---
+## xm_transformer_s2ut_800m-es-en-st-asr-bt_h1_2022
 
-w2v : /checkpoint/andyyuan/w2v2/hokkien/fbank/08-08/hok_w2v2.fsdp.TaiwaneseYT30k_30sec_train_with_TD.mt6000.317m.qtz.mlp768.ld0.0.drp_i0.0.drp_f0.0.fgm1.0.pos_enc_rope.d_0.0.pen0.2_0.mu1000000.lr0.0005.uf1.wu32000.ngpu128/checkpoint_76_900000.pt  
+Speech-to-speech translation model from fairseq S2UT ([paper](https://arxiv.org/abs/2204.02967)/[code](https://github.com/facebookresearch/fairseq/blob/main/examples/speech_to_speech/docs/enhanced_direct_s2st_discrete_units.md)):
+- Spanish-English
+- Trained on mTEDx, CoVoST 2, Europarl-ST and VoxPopuli
+- Speech synthesis with [facebook/unit_hifigan_mhubert_vp_en_es_fr_it3_400k_layer11_km1000_lj_dur](https://huggingface.co/facebook/unit_hifigan_mhubert_vp_en_es_fr_it3_400k_layer11_km1000_lj_dur)
 
-config: config_t2_fbank.yaml and mtl_config_t2_tug65k8.0_rdrop10.0.yaml under the data directory. 
+## Usage
+```python
+import json
+import os
+from pathlib import Path
 
+import IPython.display as ipd
+from fairseq import hub_utils
+from fairseq.checkpoint_utils import load_model_ensemble_and_task_from_hf_hub
+from fairseq.models.speech_to_text.hub_interface import S2THubInterface
+from fairseq.models.text_to_speech import CodeHiFiGANVocoder
+from fairseq.models.text_to_speech.hub_interface import VocoderHubInterface
 
+from huggingface_hub import snapshot_download
+import torchaudio
+
+cache_dir = os.getenv("HUGGINGFACE_HUB_CACHE")
+
+models, cfg, task = load_model_ensemble_and_task_from_hf_hub(
+    "facebook/xm_transformer_s2ut_800m-es-en-st-asr-bt_h1_2022",
+    arg_overrides={"config_yaml": "config.yaml", "task": "speech_to_text"},
+    cache_dir=cache_dir,
+)
+#model = models[0].cpu()
+#cfg["task"].cpu = True
+generator = task.build_generator([model], cfg)
+
+
+# requires 16000Hz mono channel audio
+audio, _ = torchaudio.load("/path/to/an/audio/file")
+
+sample = S2THubInterface.get_model_input(task, audio)
+unit = S2THubInterface.get_prediction(task, model, generator, sample)
+
+# speech synthesis           
+library_name = "fairseq"
+cache_dir = (
+    cache_dir or (Path.home() / ".cache" / library_name).as_posix()
+)
+cache_dir = snapshot_download(
+    f"facebook/unit_hifigan_mhubert_vp_en_es_fr_it3_400k_layer11_km1000_lj_dur", cache_dir=cache_dir, library_name=library_name
+)
+
+x = hub_utils.from_pretrained(
+    cache_dir,
+    "model.pt",
+    ".",
+    archive_map=CodeHiFiGANVocoder.hub_models(),
+    config_yaml="config.json",
+    fp16=False,
+    is_vocoder=True,
+)
+
+with open(f"{x['args']['data']}/config.json") as f:
+    vocoder_cfg = json.load(f)
+assert (
+    len(x["args"]["model_path"]) == 1
+), "Too many vocoder models in the input"
+
+vocoder = CodeHiFiGANVocoder(x["args"]["model_path"][0], vocoder_cfg)
+tts_model = VocoderHubInterface(vocoder_cfg, vocoder)
+
+tts_sample = tts_model.get_model_input(unit)
+wav, sr = tts_model.get_prediction(tts_sample)
+
+ipd.Audio(wav, rate=sr)
+```
+
+## Citation
+```bibtex
+@misc{https://doi.org/10.48550/arxiv.2204.02967,
+  doi = {10.48550/ARXIV.2204.02967},  
+  url = {https://arxiv.org/abs/2204.02967},  
+  author = {Popuri, Sravya and Chen, Peng-Jen and Wang, Changhan and Pino, Juan and Adi, Yossi and Gu, Jiatao and Hsu, Wei-Ning and Lee, Ann},  
+  keywords = {Computation and Language (cs.CL), Sound (cs.SD), Audio and Speech Processing (eess.AS), FOS: Computer and information sciences, FOS: Computer and information sciences, FOS: Electrical engineering, electronic engineering, information engineering, FOS: Electrical engineering, electronic engineering, information engineering},  
+  title = {Enhanced Direct Speech-to-Speech Translation Using Self-supervised Pre-training and Data Augmentation},  
+  publisher = {arXiv},  
+  year = {2022},  
+  copyright = {arXiv.org perpetual, non-exclusive license}
+}
+```