Update README.md

README.md

@@ -1,3 +1,250 @@
----
-license: cc-by-nc-4.0
----
+---
+license: cc-by-nc-4.0
+language:
+- de
+base_model:
+- HKUSTAudio/Llasa-1B-Multilingual
+---
+
+
+# Llasa-1B-Multilingual-German
+
+> This model was trained on top of [HKUSTAudio/Llasa-1B-Multilingual](https://huggingface.co/HKUSTAudio/Llasa-1B-Multilingual).
+
+## Model Overview
+
+This text-to-speech (TTS) model has been trained on a custom dataset representing **7,000 hours** of high-quality audio data. The audio data consisted of permissive podcasts, lectures and other OER data. 
+
+## Training Details
+
+- **Base Model:** HKUSTAudio/Llasa-1B-Multilingual
+- **Dataset:** A custom dataset comprising **7,000 hours** of data.
+- **Compute Resources:** The training was performed using **4x L40s GPUs**.
+- **Raw Training Time:** Approximately **20 hours** not included the data preprocessing with xcodec2 (note: training was restarted after 3 crashes).
+  
+Huge thanks to Hugging Face for their generous GPU grant! 🤗
+
+
+## 👨‍💻 Installation
+First install the following pip packages: 
+```bash
+pip install xcodec2
+pip install torch==2.6.0 torchaudio
+```
+Install it in the two steps given above! If you get the error message with "flex attention" make sure to install `torch==2.6.0 torchaudio`. If you get an torchaudio error, make sure to update and match it to the torch 2.6.0 version. 
+
+## 🛠️ Usage
+### 🎲 Random voice
+A basic example using the Hugging Face Transformers:
+
+```python
+import os
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import torch
+import soundfile as sf
+
+llasa_1b_german = 'MultiLlasa/Llasa-1B-Multilingual-German'  
+
+# Loading the model 
+tokenizer = AutoTokenizer.from_pretrained(llasa_1b_german)
+model = AutoModelForCausalLM.from_pretrained(llasa_1b_german)
+model.to('cuda')
+
+# Load XCodec2 model
+from xcodec2.modeling_xcodec2 import XCodec2Model
+model_path = "HKUST-Audio/xcodec2"
+Codec_model = XCodec2Model.from_pretrained(model_path)
+Codec_model.cuda()
+
+input_text = "\"Weißt du was, Hoppi\", sagte der weise Uhu, \"manchmal ist es gar nicht so wichtig, das Ende des Regenbogens zu finden. Das Schönste ist doch, dass wir alle zusammen dieses Abenteuer erleben!"
+
+
+def extract_speech_ids(speech_tokens_str):
+    speech_ids = []
+    for token_str in speech_tokens_str:
+        if token_str.startswith('<|s_') and token_str.endswith('|>'):
+            num_str = token_str[4:-2]
+            num = int(num_str)
+            speech_ids.append(num)
+        else:
+            print(f"Unexpected token: {token_str}")
+    return speech_ids
+
+with torch.no_grad():
+    formatted_text = f"<|TEXT_UNDERSTANDING_START|>{input_text}<|TEXT_UNDERSTANDING_END|>"
+    
+    chat = [
+        {"role": "user", "content": "Convert the text to speech:" + formatted_text},
+        {"role": "assistant", "content": "<|SPEECH_GENERATION_START|>"}
+    ]
+
+    input_ids = tokenizer.apply_chat_template(
+        chat,
+        tokenize=True,
+        return_tensors='pt',
+        continue_final_message=True
+    )
+    input_ids = input_ids.to('cuda')
+    speech_end_id = tokenizer.convert_tokens_to_ids('<|SPEECH_GENERATION_END|>')
+
+    outputs = model.generate(
+        input_ids,
+        max_length=2048,
+        eos_token_id=speech_end_id,
+        do_sample=True,
+        top_p=1,
+        temperature=0.8,
+    )
+
+    generated_ids = outputs[0][input_ids.shape[1]:-1]
+    speech_tokens = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+    speech_tokens = extract_speech_ids(speech_tokens)
+    speech_tokens = torch.tensor(speech_tokens).cuda().unsqueeze(0).unsqueeze(0)
+    gen_wav = Codec_model.decode_code(speech_tokens)
+    
+    
+sf.write("generation.wav", gen_wav[0, 0, :].cpu().numpy(), 16000)
+
+```
+
+### 🎯 Using a specific speaker
+
+An example with speaker reference: 
+```python 
+import torch
+import torchaudio
+import tempfile
+import soundfile as sf
+from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
+
+# Input your reference audio and optional the text
+sample_audio_path = "male.wav"
+sample_audio_text = None # Set it to none to use whisper for transcription
+# Input the target text here
+target_text = "Und apropos Spannungen und Unfälle, in Stuttgart gibt es auch einige Schlagzeilen. Die Polizei sucht Zeugen, nachdem in der Stadt mehrere Autoscheiben eingeschlagen wurden. Und gestern kam es im Stuttgarter Osten zu einer Verfolgungsjagd mit einer jungen BMW-Fahrerin, die vor einer Polizeistreife geflüchtet ist."
+output_filename = "no_speaker_example.wav"
+
+
+#### Do not edit below ####
+llasa_model_name = "MultiLlasa/Llasa-1B-Multilingual-German"
+tokenizer = AutoTokenizer.from_pretrained(llasa_model_name)
+model = AutoModelForCausalLM.from_pretrained(llasa_model_name)
+model.to("cuda")
+
+from xcodec2.modeling_xcodec2 import XCodec2Model
+codec_model_path = "HKUST-Audio/xcodec2"
+Codec_model = XCodec2Model.from_pretrained(codec_model_path)
+Codec_model.cuda()
+
+whisper_turbo_pipe = pipeline(
+    "automatic-speech-recognition",
+    model="openai/whisper-large-v3-turbo",
+    torch_dtype=torch.float16,
+    device="cuda",
+)
+
+def extract_speech_ids(speech_tokens_str_list):
+    """
+    Convert tokens like "<|s_12345|>" into integer ids.
+    """
+    speech_ids = []
+    for token_str in speech_tokens_str_list:
+        if token_str.startswith("<|s_") and token_str.endswith("|>"):
+            num_str = token_str[4:-2]
+            try:
+                speech_ids.append(int(num_str))
+            except ValueError:
+                print("Error converting token:", token_str)
+        else:
+            print(f"Unexpected token: {token_str}")
+    return speech_ids
+
+
+waveform, sample_rate = torchaudio.load(sample_audio_path)
+
+max_secs = 15
+if waveform.shape[1] / sample_rate > max_secs:
+    print("Trimming audio to the first 15 seconds.")
+    waveform = waveform[:, : sample_rate * max_secs]
+    # Pad a bit briefly (0.5 sec) at the end
+    waveform = torch.nn.functional.pad(
+        waveform, (0, int(sample_rate * 0.5)), "constant", 0
+    )
+
+if waveform.shape[0] > 1:
+    waveform = waveform.mean(dim=0, keepdim=True)
+
+if sample_rate != 16000:
+    resampler = torchaudio.transforms.Resample(orig_freq=sample_rate,
+                                                new_freq=16000)
+    waveform = resampler(waveform)
+    sample_rate = 16000
+
+if sample_audio_text is None:
+    print("Transcribing audio...")
+    transcription = whisper_turbo_pipe(waveform[0].numpy())["text"].strip()
+else:
+    transcription = sample_audio_text
+
+print("Transcription:", transcription)
+
+if len(target_text) == 0:
+    raise ValueError("Target text must be provided!")
+elif len(target_text) > 500:
+    print("Text is too long; trimming to first 500 characters.")
+    target_text = target_text[:500]
+
+input_text = transcription + " " + target_text
+
+formatted_text = f"<|TEXT_UNDERSTANDING_START|>{input_text}<|TEXT_UNDERSTANDING_END|>"
+
+chat = [
+    {"role": "user", "content": "Convert the text to speech:" + formatted_text},
+    {"role": "assistant", "content": "<|SPEECH_GENERATION_START|>"}
+]
+
+input_ids = tokenizer.apply_chat_template(
+    chat, tokenize=True, return_tensors="pt", continue_final_message=True
+)
+input_ids = input_ids.to("cuda")
+speech_end_id = tokenizer.convert_tokens_to_ids("<|SPEECH_GENERATION_END|>")
+
+with torch.no_grad():
+    outputs = model.generate(
+        input_ids,
+        max_length=2048,
+        eos_token_id=speech_end_id,
+        do_sample=True,
+        top_p=1,
+        temperature=0.8,
+    )
+
+    generated_ids = outputs[0][input_ids.shape[1] : -1]
+
+    raw_speech_tokens = tokenizer.batch_decode(generated_ids,
+                                                skip_special_tokens=True)
+    speech_ids = extract_speech_ids(raw_speech_tokens)
+
+    if len(speech_ids) == 0:
+        raise ValueError("No valid speech tokens were generated!")
+
+    speech_tokens_tensor = torch.tensor(speech_ids)\
+        .cuda().unsqueeze(0).unsqueeze(0)
+
+    gen_wav = Codec_model.decode_code(speech_tokens_tensor).cpu().squeeze()
+
+sf.write(output_filename, gen_wav, 16000)
+```
+
+
+## Tips
+- With a reference speaker, audio glitches can happen. Try to increase the temperature to get better results. 
+
+## License
+
+This project is licensed under the [CC-BY-NC-4.0 license](https://creativecommons.org/licenses/by-nc/4.0/).
+
+## Acknowledgments
+
+- **Hugging Face:** Thanks for the grant that made this project possible.
+