Add example scripts for monologue TTS.

cli/tts.py

@@ -0,0 +1,83 @@
+import os
+import json
+import torch
+import argparse
+
+import s3tokenizer
+import soundfile as sf
+
+from soulxpodcast.config import SamplingParams
+from soulxpodcast.utils.parser import podcast_format_parser
+from soulxpodcast.utils.infer_utils import initiate_model, process_single_input
+
+
+def run_inference(
+    inputs: dict,
+    model_path: str,
+    output_path: str,
+    llm_engine: str = "hf",
+    fp16_flow: bool = False,
+    seed: int = 1988,
+):
+    
+    model, dataset = initiate_model(seed, model_path, llm_engine, fp16_flow)
+    
+    data = process_single_input(
+        dataset,
+        inputs['text'],
+        inputs['prompt_wav'],
+        inputs['prompt_text'],
+        inputs['use_dialect_prompt'],
+        inputs['dialect_prompt_text'],
+    )
+
+    print("[INFO] Start inference...")
+    results_dict = model.forward_longform(**data)
+
+    target_audio = None
+    for wav in results_dict["generated_wavs"]:
+        if target_audio is None:
+            target_audio = wav
+        else:
+            target_audio = torch.cat([target_audio, wav], dim=1)
+
+    os.makedirs(os.path.dirname(output_path), exist_ok=True)
+    sf.write(output_path, target_audio.cpu().squeeze(0).numpy(), 24000)
+    print(f"[INFO] Saved synthesized audio to: {output_path}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--text", required=True, help="The text used for generating real audio")
+    parser.add_argument("--prompt_text", required=True, help="The text used for prompt")
+    parser.add_argument("--dialect_prompt", default="", help="The prompt dialect text used for prompt")
+    parser.add_argument("--prompt_audio", required=True, help="Path to the input JSON file")
+    parser.add_argument("--model_path", required=True, help="Path to the model file")
+    parser.add_argument("--output_path", default="outputs/result.wav", help="Path to the output audio file")
+    parser.add_argument("--llm_engine", default="hf", choices=["hf", "vllm"], help="Inference engine to use")
+    parser.add_argument("--fp16_flow", action="store_true", help="Enable FP16 flow")
+    parser.add_argument("--seed", type=int, default=1988, help="Random seed")
+    args = parser.parse_args()
+
+    data = {
+        "speakers":{
+            "S1":{
+                "prompt_audio": args.prompt_audio,
+                "prompt_text": args.prompt_text,
+                "dialect_prompt": args.dialect_prompt,
+            }
+        },
+        "text": [
+            ["S1", args.text]
+        ]
+    }
+    inputs = podcast_format_parser(data)
+    run_inference(
+        inputs=inputs,
+        model_path=args.model_path,
+        output_path=args.output_path,
+        llm_engine=args.llm_engine,
+        fp16_flow=args.fp16_flow,
+        seed=args.seed,
+    )

example/infer_tts.sh

@@ -0,0 +1,24 @@
+export PYTHONPATH="$(cd "$(dirname "${BASH_SOURCE[0]}")"/.. && pwd)"
+echo "PYTHONPATH set to: $PYTHONPATH"
+
+# Note: To infer Chinese dialects, set model_dir to "pretrained_models/SoulX-Podcast-1.7B-dialect 
+#       and follow example/podcast_script/script_mandarin to set dialect_prompt and pass it to 
+#       entrance tts.py.
+# Example: infer chinese sichuan dialects with mandarin prompt:
+# dialect_prompt="<|Sichuan|>要得要得！前头几个耍洋盘，我后脚就背起铺盖卷去景德镇耍泥巴，巴适得喊老天爷！"
+# text="<|Sichuan|>一个着迷于在岩壁与雪原间捕捉语言灵感的旅人，即将奔赴景德镇将朝露炊烟揉进陶胚的造梦者。"
+
+model_dir=pretrained_models/SoulX-Podcast-1.7B
+prompt_text="喜欢攀岩、徒步、滑雪的语言爱好者，以及过两天要带着全部家当去景德镇做陶瓷的白日梦想家。"
+dialect_prompt=""
+prompt_audio="example/audios/female_mandarin.wav"
+text="一个着迷于在岩壁与雪原间捕捉语言灵感的旅人，即将奔赴景德镇将朝露炊烟揉进陶胚的造梦者。"
+
+python cli/tts.py \
+        --prompt_text ${prompt_text} \
+        --dialect_prompt "${dialect_prompt:-}"  \
+        --prompt_audio ${prompt_audio} \
+        --text ${text} \
+        --model_path ${model_dir} \
+        --output_path outputs/mandarin_tts.wav \
+        --seed 7