Spaces:
Bradarr
/
Running on Zero

Zackh commited on
Commit
96dc011
·
1 Parent(s): fa0bc40
.gitattributes CHANGED
@@ -33,3 +33,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ prompts/conversational_a.wav filter=lfs diff=lfs merge=lfs -text
37
+ prompts/conversational_b.wav filter=lfs diff=lfs merge=lfs -text
38
+ prompts/read_speech_a.wav filter=lfs diff=lfs merge=lfs -text
39
+ prompts/read_speech_b.wav filter=lfs diff=lfs merge=lfs -text
40
+ prompts/read_speech_c.wav filter=lfs diff=lfs merge=lfs -text
41
+ prompts/read_speech_d.wav filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -1,5 +1,5 @@
1
  ---
2
- title: CSM 1B
3
  emoji: 🚀
4
  colorFrom: blue
5
  colorTo: blue
@@ -8,7 +8,36 @@ sdk_version: 5.20.0
8
  app_file: app.py
9
  pinned: false
10
  license: apache-2.0
11
- short_description: 1B variant of Sesame's Conversational Speech Model
12
  ---
13
 
14
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ title: Sesame CSM Space
3
  emoji: 🚀
4
  colorFrom: blue
5
  colorTo: blue
 
8
  app_file: app.py
9
  pinned: false
10
  license: apache-2.0
11
+ short_description: Generation using Sesame's Conversational Speech Model
12
  ---
13
 
14
+ ## CSM 1B
15
+
16
+ **2025/03/13** - We are releasing the 1B CSM variant. Code is available on GitHub: [SesameAILabs/csm](https://github.com/SesameAILabs/csm). Checkpoint is [hosted on HuggingFace](https://huggingface.co/sesame/csm-1b).
17
+
18
+ Try out the interactive demo of our fine-tuned version [sesame.com/voicedemo](https://www.sesame.com/voicedemo).
19
+
20
+ Generate from the open-source base model [hosted on HuggingFace](https://huggingface.co/spaces/sesame/csm-1b).
21
+
22
+ ---
23
+
24
+ CSM (Conversational Speech Model) is a speech generation model from [Sesame](sesame.com) that generates RVQ audio codes from text and audio inputs. A fine-tuned version of this model powers the interactive demo in our [technical blog post](https://www.sesame.com/research/crossing_the_uncanny_valley_of_voice).
25
+
26
+ The model architecture employs a [Llama](https://www.llama.com/) backbone and a smaller audio decoder that produces [Mimi](https://huggingface.co/kyutai/mimi) audio codes.
27
+
28
+ ## Misuse and abuse ⚠️
29
+
30
+ This project provides a high-quality speech generation model for research and educational purposes. While we encourage responsible and ethical use, we **explicitly prohibit** the following:
31
+
32
+ - **Impersonation or Fraud**: Do not use this model to generate speech that mimics real individuals without their explicit consent.
33
+ - **Misinformation or Deception**: Do not use this model to create deceptive or misleading content, such as fake news or fraudulent calls.
34
+ - **Illegal or Harmful Activities**: Do not use this model for any illegal, harmful, or malicious purposes.
35
+
36
+ By using this model, you agree to comply with all applicable laws and ethical guidelines. We are **not responsible** for any misuse, and we strongly condemn unethical applications of this technology.
37
+
38
+ **Prompts**
39
+ Conversational prompts are from the [EdAcc dataset](https://groups.inf.ed.ac.uk/edacc/)
40
+ Read speech prompts are form the [LibriTTS-R dataset](https://google.github.io/df-conformer/librittsr/)
41
+
42
+ **Authors**
43
+ Johan Schalkwyk, Ankit Kumar, Dan Lyth, Sefik Emre Eskimez, Zack Hodari, Cinjon Resnick, Ramon Sanabria, Raven Jiang, and the Sesame team.
app.py ADDED
@@ -0,0 +1,232 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+
3
+ import gradio as gr
4
+ import numpy as np
5
+ import spaces
6
+ import torch
7
+ import torchaudio
8
+ from generator import Segment, load_csm_1b
9
+ from huggingface_hub import hf_hub_download, login
10
+ from watermarking import watermark
11
+
12
+ api_key = os.getenv("HF_TOKEN")
13
+ gpu_timeout = int(os.getenv("GPU_TIMEOUT", 60))
14
+ CSM_1B_HF_WATERMARK = list(map(int, os.getenv("WATERMARK_KEY").split(" ")))
15
+
16
+ login(token=api_key)
17
+
18
+ SPACE_INTRO_TEXT = """\
19
+ # Sesame CSM 1B
20
+
21
+ Generate from CSM 1B (Conversational Speech Model).
22
+ Code is available on GitHub: [SesameAILabs/csm](https://github.com/SesameAILabs/csm).
23
+ Checkpoint is [hosted on HuggingFace](https://huggingface.co/sesame/csm-1b).
24
+
25
+ Try out the interactive demo of our fine-tuned model [sesame.com/voicedemo](https://www.sesame.com/voicedemo).
26
+
27
+ The model has some capacity for non-English languages due to data contamination in the training
28
+ data, but it is likely not to perform well.
29
+
30
+ ---
31
+
32
+ """
33
+
34
+ CONVO_INTRO_TEXT = """\
35
+ ## Conversation content
36
+
37
+ Each line is an utterance in the conversation to generate. Speakers alternate between A and B, starting with speaker A.
38
+ """
39
+
40
+ DEFAULT_CONVERSATION = """\
41
+ Hey how are you doing.
42
+ Pretty good, pretty good.
43
+ I'm great, so happy to be speaking to you.
44
+ Me too, this is some cool stuff huh?
45
+ Yeah, I've been reading more about speech generation, and it really seems like context is important.
46
+ Definitely.
47
+ """
48
+
49
+ SPEAKER_PROMPTS = {
50
+ "conversational_a": {
51
+ "text": (
52
+ "like revising for an exam I'd have to try and like keep up the momentum because I'd "
53
+ "start really early I'd be like okay I'm gonna start revising now and then like "
54
+ "you're revising for ages and then I just like start losing steam I didn't do that "
55
+ "for the exam we had recently to be fair that was a more of a last minute scenario "
56
+ "but like yeah I'm trying to like yeah I noticed this yesterday that like Mondays I "
57
+ "sort of start the day with this not like a panic but like a"
58
+ ),
59
+ "audio": "prompts/conversational_a.wav",
60
+ },
61
+ "conversational_b": {
62
+ "text": (
63
+ "like a super Mario level. Like it's very like high detail. And like, once you get "
64
+ "into the park, it just like, everything looks like a computer game and they have all "
65
+ "these, like, you know, if, if there's like a, you know, like in a Mario game, they "
66
+ "will have like a question block. And if you like, you know, punch it, a coin will "
67
+ "come out. So like everyone, when they come into the park, they get like this little "
68
+ "bracelet and then you can go punching question blocks around."
69
+ ),
70
+ "audio": "prompts/conversational_b.wav",
71
+ },
72
+ "read_speech_a": {
73
+ "text": (
74
+ "And Lake turned round upon me, a little abruptly, his odd yellowish eyes, a little "
75
+ "like those of the sea eagle, and the ghost of his smile that flickered on his "
76
+ "singularly pale face, with a stern and insidious look, confronted me."
77
+ ),
78
+ "audio": "prompts/read_speech_a.wav",
79
+ },
80
+ "read_speech_b": {
81
+ "text": (
82
+ "He was such a big boy that he wore high boots and carried a jack knife. He gazed and "
83
+ "gazed at the cap, and could not keep from fingering the blue tassel."
84
+ ),
85
+ "audio": "prompts/read_speech_b.wav",
86
+ },
87
+ "read_speech_c": {
88
+ "text": (
89
+ "All passed so quickly, there was so much going on around him, the Tree quite forgot "
90
+ "to look to himself."
91
+ ),
92
+ "audio": "prompts/read_speech_c.wav",
93
+ },
94
+ "read_speech_d": {
95
+ "text": (
96
+ "Suddenly I was back in the old days Before you felt we ought to drift apart. It was "
97
+ "some trick-the way your eyebrows raise."
98
+ ),
99
+ "audio": "prompts/read_speech_d.wav",
100
+ },
101
+ }
102
+
103
+ device = "cuda" if torch.cuda.is_available() else "cpu"
104
+ model_path = hf_hub_download(repo_id="sesame/csm-1b", filename="ckpt.pt")
105
+ generator = load_csm_1b(model_path, device)
106
+
107
+
108
+ @spaces.GPU(duration=gpu_timeout)
109
+ def infer(
110
+ text_prompt_speaker_a,
111
+ text_prompt_speaker_b,
112
+ audio_prompt_speaker_a,
113
+ audio_prompt_speaker_b,
114
+ gen_conversation_input,
115
+ ) -> tuple[np.ndarray, int]:
116
+ audio_prompt_a = prepare_prompt(text_prompt_speaker_a, 0, audio_prompt_speaker_a)
117
+ audio_prompt_b = prepare_prompt(text_prompt_speaker_b, 1, audio_prompt_speaker_b)
118
+
119
+ prompt_segments: list[Segment] = [audio_prompt_a, audio_prompt_b]
120
+ generated_segments: list[Segment] = []
121
+
122
+ conversation_lines = [line.strip() for line in gen_conversation_input.strip().split("\n") if line.strip()]
123
+ for i, line in enumerate(conversation_lines):
124
+ # Alternating speakers A and B, starting with A
125
+ speaker_id = i % 2
126
+
127
+ audio_tensor = generator.generate(
128
+ text=line,
129
+ speaker=speaker_id,
130
+ context=prompt_segments + generated_segments,
131
+ )
132
+ generated_segments.append(Segment(text=line, speaker=speaker_id, audio=audio_tensor))
133
+
134
+ # Concatenate all generations and convert to 16-bit int format
135
+ audio_tensors = [segment.audio for segment in generated_segments]
136
+ audio_tensor = torch.cat(audio_tensors, dim=0)
137
+
138
+ # This applies an imperceptible watermark to identify audio as AI-generated.
139
+ # Watermarking ensures transparency, dissuades misuse, and enables traceability.
140
+ # Please be a responsible AI citizen and keep the watermarking in place.
141
+ # If using CSM 1B in another application, use your own private key and keep it secret.
142
+ audio_tensor, wm_sample_rate = watermark(
143
+ generator._watermarker, audio_tensor, generator.sample_rate, CSM_1B_HF_WATERMARK
144
+ )
145
+ audio_tensor = torchaudio.functional.resample(
146
+ audio_tensor, orig_freq=wm_sample_rate, new_freq=generator.sample_rate
147
+ )
148
+
149
+ audio_array = (audio_tensor * 32768).to(torch.int16).cpu().numpy()
150
+
151
+ return generator.sample_rate, audio_array
152
+
153
+
154
+ def prepare_prompt(text: str, speaker: int, audio_path: str) -> Segment:
155
+ audio_tensor, _ = load_prompt_audio(audio_path)
156
+ return Segment(text=text, speaker=speaker, audio=audio_tensor)
157
+
158
+
159
+ def load_prompt_audio(audio_path: str) -> torch.Tensor:
160
+ audio_tensor, sample_rate = torchaudio.load(audio_path)
161
+ audio_tensor = audio_tensor.squeeze(0)
162
+ if sample_rate != generator.sample_rate:
163
+ audio_tensor = torchaudio.functional.resample(
164
+ audio_tensor, orig_freq=sample_rate, new_freq=generator.sample_rate
165
+ )
166
+ return audio_tensor, generator.sample_rate
167
+
168
+
169
+ def create_speaker_prompt_ui(speaker_name: str):
170
+ speaker_dropdown = gr.Dropdown(
171
+ choices=list(SPEAKER_PROMPTS.keys()), label="Select a predefined speaker", value=speaker_name
172
+ )
173
+ with gr.Accordion("Or add your own voice prompt", open=False):
174
+ text_prompt_speaker = gr.Textbox(label="Speaker prompt", lines=4, value=SPEAKER_PROMPTS[speaker_name]["text"])
175
+ audio_prompt_speaker = gr.Audio(
176
+ label="Speaker prompt", type="filepath", value=SPEAKER_PROMPTS[speaker_name]["audio"]
177
+ )
178
+
179
+ return speaker_dropdown, text_prompt_speaker, audio_prompt_speaker
180
+
181
+
182
+ with gr.Blocks() as app:
183
+ gr.Markdown(SPACE_INTRO_TEXT)
184
+ gr.Markdown("## Voices")
185
+ with gr.Row():
186
+ with gr.Column():
187
+ gr.Markdown("### Speaker A")
188
+ speaker_a_dropdown, text_prompt_speaker_a, audio_prompt_speaker_a = create_speaker_prompt_ui(
189
+ "conversational_a"
190
+ )
191
+
192
+ with gr.Column():
193
+ gr.Markdown("### Speaker B")
194
+ speaker_b_dropdown, text_prompt_speaker_b, audio_prompt_speaker_b = create_speaker_prompt_ui(
195
+ "conversational_b"
196
+ )
197
+
198
+ def update_audio(speaker):
199
+ if speaker in SPEAKER_PROMPTS:
200
+ return SPEAKER_PROMPTS[speaker]["audio"]
201
+ return None
202
+
203
+ def update_text(speaker):
204
+ if speaker in SPEAKER_PROMPTS:
205
+ return SPEAKER_PROMPTS[speaker]["text"]
206
+ return None
207
+
208
+ speaker_a_dropdown.change(fn=update_audio, inputs=[speaker_a_dropdown], outputs=[audio_prompt_speaker_a])
209
+ speaker_b_dropdown.change(fn=update_audio, inputs=[speaker_b_dropdown], outputs=[audio_prompt_speaker_b])
210
+
211
+ speaker_a_dropdown.change(fn=update_text, inputs=[speaker_a_dropdown], outputs=[text_prompt_speaker_a])
212
+ speaker_b_dropdown.change(fn=update_text, inputs=[speaker_b_dropdown], outputs=[text_prompt_speaker_b])
213
+
214
+ gr.Markdown(CONVO_INTRO_TEXT)
215
+
216
+ gen_conversation_input = gr.TextArea(label="conversation", lines=20, value=DEFAULT_CONVERSATION)
217
+ generate_btn = gr.Button("Generate conversation", variant="primary")
218
+ audio_output = gr.Audio(label="Synthesized audio")
219
+
220
+ generate_btn.click(
221
+ infer,
222
+ inputs=[
223
+ text_prompt_speaker_a,
224
+ text_prompt_speaker_b,
225
+ audio_prompt_speaker_a,
226
+ audio_prompt_speaker_b,
227
+ gen_conversation_input,
228
+ ],
229
+ outputs=[audio_output],
230
+ )
231
+
232
+ app.launch(ssr_mode=True)
generator.py ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ from dataclasses import dataclass
3
+ from typing import List, Tuple
4
+
5
+ import torch
6
+ import torchaudio
7
+ from huggingface_hub import hf_hub_download
8
+ from models import Model, ModelArgs
9
+ from moshi.models import loaders
10
+ from tokenizers.processors import TemplateProcessing
11
+ from transformers import AutoTokenizer
12
+ from watermarking import load_watermarker, watermark
13
+
14
+ CSM_1B_HF_WATERMARK = list(map(int, os.getenv("WATERMARK_KEY").split(" ")))
15
+
16
+
17
+ @dataclass
18
+ class Segment:
19
+ speaker: int
20
+ text: str
21
+ # (num_samples,), sample_rate = 24_000
22
+ audio: torch.Tensor
23
+
24
+
25
+ def load_llama3_tokenizer():
26
+ """
27
+ https://github.com/huggingface/transformers/issues/22794#issuecomment-2092623992
28
+ """
29
+ tokenizer_name = "meta-llama/Llama-3.2-1B"
30
+ tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
31
+ bos = tokenizer.bos_token
32
+ eos = tokenizer.eos_token
33
+ tokenizer._tokenizer.post_processor = TemplateProcessing(
34
+ single=f"{bos}:0 $A:0 {eos}:0",
35
+ pair=f"{bos}:0 $A:0 {eos}:0 {bos}:1 $B:1 {eos}:1",
36
+ special_tokens=[(f"{bos}", tokenizer.bos_token_id), (f"{eos}", tokenizer.eos_token_id)],
37
+ )
38
+
39
+ return tokenizer
40
+
41
+
42
+ class Generator:
43
+ def __init__(
44
+ self,
45
+ model: Model,
46
+ ):
47
+ self._model = model
48
+ self._model.setup_caches(1)
49
+
50
+ self._text_tokenizer = load_llama3_tokenizer()
51
+
52
+ device = next(model.parameters()).device
53
+ mimi_weight = hf_hub_download(loaders.DEFAULT_REPO, loaders.MIMI_NAME)
54
+ mimi = loaders.get_mimi(mimi_weight, device=device)
55
+ mimi.set_num_codebooks(32)
56
+ self._audio_tokenizer = mimi
57
+
58
+ self._watermarker = load_watermarker(device=device)
59
+
60
+ self.sample_rate = mimi.sample_rate
61
+ self.device = device
62
+
63
+ def _tokenize_text_segment(self, text: str, speaker: int) -> Tuple[torch.Tensor, torch.Tensor]:
64
+ frame_tokens = []
65
+ frame_masks = []
66
+
67
+ text_tokens = self._text_tokenizer.encode(f"[{speaker}]{text}")
68
+ text_frame = torch.zeros(len(text_tokens), 33).long()
69
+ text_frame_mask = torch.zeros(len(text_tokens), 33).bool()
70
+ text_frame[:, -1] = torch.tensor(text_tokens)
71
+ text_frame_mask[:, -1] = True
72
+
73
+ frame_tokens.append(text_frame.to(self.device))
74
+ frame_masks.append(text_frame_mask.to(self.device))
75
+
76
+ return torch.cat(frame_tokens, dim=0), torch.cat(frame_masks, dim=0)
77
+
78
+ def _tokenize_audio(self, audio: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
79
+ frame_tokens = []
80
+ frame_masks = []
81
+
82
+ # (K, T)
83
+ audio = audio.to(self.device)
84
+ audio_tokens = self._audio_tokenizer.encode(audio.unsqueeze(0).unsqueeze(0))[0]
85
+ # add EOS frame
86
+ eos_frame = torch.zeros(audio_tokens.size(0), 1).to(self.device)
87
+ audio_tokens = torch.cat([audio_tokens, eos_frame], dim=1)
88
+
89
+ audio_frame = torch.zeros(audio_tokens.size(1), 33).long().to(self.device)
90
+ audio_frame_mask = torch.zeros(audio_tokens.size(1), 33).bool().to(self.device)
91
+ audio_frame[:, :-1] = audio_tokens.transpose(0, 1)
92
+ audio_frame_mask[:, :-1] = True
93
+
94
+ frame_tokens.append(audio_frame)
95
+ frame_masks.append(audio_frame_mask)
96
+
97
+ return torch.cat(frame_tokens, dim=0), torch.cat(frame_masks, dim=0)
98
+
99
+ def _tokenize_segment(self, segment: Segment) -> Tuple[torch.Tensor, torch.Tensor]:
100
+ """
101
+ Returns:
102
+ (seq_len, 33), (seq_len, 33)
103
+ """
104
+ text_tokens, text_masks = self._tokenize_text_segment(segment.text, segment.speaker)
105
+ audio_tokens, audio_masks = self._tokenize_audio(segment.audio)
106
+
107
+ return torch.cat([text_tokens, audio_tokens], dim=0), torch.cat([text_masks, audio_masks], dim=0)
108
+
109
+ @torch.inference_mode()
110
+ def generate(
111
+ self,
112
+ text: str,
113
+ speaker: int,
114
+ context: List[Segment],
115
+ max_audio_length_ms: float = 90_000,
116
+ temperature: float = 0.9,
117
+ topk: int = 50,
118
+ ) -> torch.Tensor:
119
+ self._model.reset_caches()
120
+
121
+ max_audio_frames = int(max_audio_length_ms / 80)
122
+ tokens, tokens_mask = [], []
123
+ for segment in context:
124
+ segment_tokens, segment_tokens_mask = self._tokenize_segment(segment)
125
+ tokens.append(segment_tokens)
126
+ tokens_mask.append(segment_tokens_mask)
127
+
128
+ gen_segment_tokens, gen_segment_tokens_mask = self._tokenize_text_segment(text, speaker)
129
+ tokens.append(gen_segment_tokens)
130
+ tokens_mask.append(gen_segment_tokens_mask)
131
+
132
+ prompt_tokens = torch.cat(tokens, dim=0).long().to(self.device)
133
+ prompt_tokens_mask = torch.cat(tokens_mask, dim=0).bool().to(self.device)
134
+
135
+ samples = []
136
+ curr_tokens = prompt_tokens.unsqueeze(0)
137
+ curr_tokens_mask = prompt_tokens_mask.unsqueeze(0)
138
+ curr_pos = torch.arange(0, prompt_tokens.size(0)).unsqueeze(0).long().to(self.device)
139
+
140
+ for _ in range(max_audio_frames):
141
+ sample = self._model.generate_frame(curr_tokens, curr_tokens_mask, curr_pos, temperature, topk)
142
+ if torch.all(sample == 0):
143
+ break # eos
144
+
145
+ samples.append(sample)
146
+
147
+ curr_tokens = torch.cat([sample, torch.zeros(1, 1).long().to(self.device)], dim=1).unsqueeze(1)
148
+ curr_tokens_mask = torch.cat(
149
+ [torch.ones_like(sample).bool(), torch.zeros(1, 1).bool().to(self.device)], dim=1
150
+ ).unsqueeze(1)
151
+ curr_pos = curr_pos[:, -1:] + 1
152
+
153
+ audio = self._audio_tokenizer.decode(torch.stack(samples).permute(1, 2, 0)).squeeze(0).squeeze(0)
154
+
155
+ # This applies an imperceptible watermark to identify audio as AI-generated.
156
+ # Watermarking ensures transparency, dissuades misuse, and enables traceability.
157
+ # Please be a responsible AI citizen and keep the watermarking in place.
158
+ # If using CSM 1B in another application, use your own private key and keep it secret.
159
+ audio, wm_sample_rate = watermark(self._watermarker, audio, self.sample_rate, CSM_1B_HF_WATERMARK)
160
+ audio = torchaudio.functional.resample(audio, orig_freq=wm_sample_rate, new_freq=self.sample_rate)
161
+
162
+ return audio
163
+
164
+
165
+ def load_csm_1b(ckpt_path: str = "ckpt.pt", device: str = "cuda") -> Generator:
166
+ model_args = ModelArgs(
167
+ backbone_flavor="llama-1B",
168
+ decoder_flavor="llama-100M",
169
+ text_vocab_size=128256,
170
+ audio_vocab_size=2051,
171
+ audio_num_codebooks=32,
172
+ )
173
+ model = Model(model_args).to(device=device, dtype=torch.bfloat16)
174
+ state_dict = torch.load(ckpt_path)
175
+ model.load_state_dict(state_dict)
176
+
177
+ generator = Generator(model)
178
+ return generator
models.py ADDED
@@ -0,0 +1,196 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from dataclasses import dataclass
2
+
3
+ import torch
4
+ import torch.nn as nn
5
+ import torchtune
6
+ from torchtune.models import llama3_2
7
+
8
+
9
+ def llama3_2_1B() -> torchtune.modules.transformer.TransformerDecoder:
10
+ return llama3_2.llama3_2(
11
+ vocab_size=128_256,
12
+ num_layers=16,
13
+ num_heads=32,
14
+ num_kv_heads=8,
15
+ embed_dim=2048,
16
+ max_seq_len=2048,
17
+ intermediate_dim=8192,
18
+ attn_dropout=0.0,
19
+ norm_eps=1e-5,
20
+ rope_base=500_000,
21
+ scale_factor=32,
22
+ )
23
+
24
+
25
+ def llama3_2_100M() -> torchtune.modules.transformer.TransformerDecoder:
26
+ return llama3_2.llama3_2(
27
+ vocab_size=128_256,
28
+ num_layers=4,
29
+ num_heads=8,
30
+ num_kv_heads=2,
31
+ embed_dim=1024,
32
+ max_seq_len=2048,
33
+ intermediate_dim=8192,
34
+ attn_dropout=0.0,
35
+ norm_eps=1e-5,
36
+ rope_base=500_000,
37
+ scale_factor=32,
38
+ )
39
+
40
+
41
+ FLAVORS = {
42
+ "llama-1B": llama3_2_1B,
43
+ "llama-100M": llama3_2_100M,
44
+ }
45
+
46
+
47
+ def _prepare_transformer(model):
48
+ embed_dim = model.tok_embeddings.embedding_dim
49
+ model.tok_embeddings = nn.Identity()
50
+ model.output = nn.Identity()
51
+ return model, embed_dim
52
+
53
+
54
+ def _create_causal_mask(seq_len: int, device: torch.device):
55
+ return torch.tril(torch.ones(seq_len, seq_len, dtype=torch.bool, device=device))
56
+
57
+
58
+ def _index_causal_mask(mask: torch.Tensor, input_pos: torch.Tensor):
59
+ """
60
+ Args:
61
+ mask: (max_seq_len, max_seq_len)
62
+ input_pos: (batch_size, seq_len)
63
+
64
+ Returns:
65
+ (batch_size, seq_len, max_seq_len)
66
+ """
67
+ r = mask[input_pos, :]
68
+ return r
69
+
70
+
71
+ def _multinomial_sample_one_no_sync(probs): # Does multinomial sampling without a cuda synchronization
72
+ q = torch.empty_like(probs).exponential_(1)
73
+ return torch.argmax(probs / q, dim=-1, keepdim=True).to(dtype=torch.int)
74
+
75
+
76
+ def sample_topk(logits: torch.Tensor, topk: int, temperature: float):
77
+ logits = logits / temperature
78
+
79
+ filter_value: float = -float("Inf")
80
+ indices_to_remove = logits < torch.topk(logits, topk)[0][..., -1, None]
81
+ scores_processed = logits.masked_fill(indices_to_remove, filter_value)
82
+ scores_processed = torch.nn.functional.log_softmax(scores_processed, dim=-1)
83
+ probs = torch.nn.functional.softmax(scores_processed, dim=-1)
84
+
85
+ sample_token = _multinomial_sample_one_no_sync(probs)
86
+ return sample_token
87
+
88
+
89
+ @dataclass
90
+ class ModelArgs:
91
+ backbone_flavor: str
92
+ decoder_flavor: str
93
+ text_vocab_size: int
94
+ audio_vocab_size: int
95
+ audio_num_codebooks: int
96
+
97
+
98
+ class Model(nn.Module):
99
+ def __init__(self, args: ModelArgs):
100
+ super().__init__()
101
+ self.args = args
102
+
103
+ self.backbone, backbone_dim = _prepare_transformer(FLAVORS[args.backbone_flavor]())
104
+ self.decoder, decoder_dim = _prepare_transformer(FLAVORS[args.decoder_flavor]())
105
+
106
+ self.text_embeddings = nn.Embedding(args.text_vocab_size, backbone_dim)
107
+ self.audio_embeddings = nn.Embedding(args.audio_vocab_size * args.audio_num_codebooks, backbone_dim)
108
+
109
+ self.projection = nn.Linear(backbone_dim, decoder_dim, bias=False)
110
+ self.codebook0_head = nn.Linear(backbone_dim, args.audio_vocab_size, bias=False)
111
+ self.audio_head = nn.Parameter(torch.empty(args.audio_num_codebooks - 1, decoder_dim, args.audio_vocab_size))
112
+
113
+ def setup_caches(self, max_batch_size: int) -> torch.Tensor:
114
+ """Setup KV caches and return a causal mask."""
115
+ dtype = next(self.parameters()).dtype
116
+ device = next(self.parameters()).device
117
+
118
+ with device:
119
+ self.backbone.setup_caches(max_batch_size, dtype)
120
+ self.decoder.setup_caches(max_batch_size, dtype, decoder_max_seq_len=self.args.audio_num_codebooks)
121
+
122
+ self.register_buffer("backbone_causal_mask", _create_causal_mask(self.backbone.max_seq_len, device))
123
+ self.register_buffer("decoder_causal_mask", _create_causal_mask(self.args.audio_num_codebooks, device))
124
+
125
+ def generate_frame(
126
+ self,
127
+ tokens: torch.Tensor,
128
+ tokens_mask: torch.Tensor,
129
+ input_pos: torch.Tensor,
130
+ temperature: float,
131
+ topk: int,
132
+ ) -> torch.Tensor:
133
+ """
134
+ Args:
135
+ tokens: (batch_size, seq_len, audio_num_codebooks+1)
136
+ tokens_mask: (batch_size, seq_len, audio_num_codebooks+1)
137
+ input_pos: (batch_size, seq_len) positions for each token
138
+ mask: (batch_size, seq_len, max_seq_len
139
+
140
+ Returns:
141
+ (batch_size, audio_num_codebooks) sampled tokens
142
+ """
143
+ dtype = next(self.parameters()).dtype
144
+ b, s, _ = tokens.size()
145
+
146
+ assert self.backbone.caches_are_enabled(), "backbone caches are not enabled"
147
+ curr_backbone_mask = _index_causal_mask(self.backbone_causal_mask, input_pos)
148
+ embeds = self._embed_tokens(tokens)
149
+ masked_embeds = embeds * tokens_mask.unsqueeze(-1)
150
+ h = masked_embeds.sum(dim=2)
151
+ h = self.backbone(h, input_pos=input_pos, mask=curr_backbone_mask).to(dtype=dtype)
152
+
153
+ last_h = h[:, -1, :]
154
+ c0_logits = self.codebook0_head(last_h)
155
+ c0_sample = sample_topk(c0_logits, topk, temperature)
156
+ c0_embed = self._embed_audio(0, c0_sample)
157
+
158
+ curr_h = torch.cat([last_h.unsqueeze(1), c0_embed], dim=1)
159
+ curr_sample = c0_sample.clone()
160
+ curr_pos = torch.arange(0, curr_h.size(1), device=curr_h.device).unsqueeze(0).repeat(curr_h.size(0), 1)
161
+
162
+ # Decoder caches must be reset every frame.
163
+ self.decoder.reset_caches()
164
+ for i in range(1, self.args.audio_num_codebooks):
165
+ curr_decoder_mask = _index_causal_mask(self.decoder_causal_mask, curr_pos)
166
+ decoder_h = self.decoder(self.projection(curr_h), input_pos=curr_pos, mask=curr_decoder_mask).to(
167
+ dtype=dtype
168
+ )
169
+ ci_logits = torch.mm(decoder_h[:, -1, :], self.audio_head[i - 1])
170
+ ci_sample = sample_topk(ci_logits, topk, temperature)
171
+ ci_embed = self._embed_audio(i, ci_sample)
172
+
173
+ curr_h = ci_embed
174
+ curr_sample = torch.cat([curr_sample, ci_sample], dim=1)
175
+ curr_pos = curr_pos[:, -1:] + 1
176
+
177
+ return curr_sample
178
+
179
+ def reset_caches(self):
180
+ self.backbone.reset_caches()
181
+ self.decoder.reset_caches()
182
+
183
+ def _embed_audio(self, codebook: int, tokens: torch.Tensor) -> torch.Tensor:
184
+ return self.audio_embeddings(tokens + codebook * self.args.audio_vocab_size)
185
+
186
+ def _embed_tokens(self, tokens: torch.Tensor) -> torch.Tensor:
187
+ text_embeds = self.text_embeddings(tokens[:, :, -1]).unsqueeze(-2)
188
+
189
+ audio_tokens = tokens[:, :, :-1] + (
190
+ self.args.audio_vocab_size * torch.arange(self.args.audio_num_codebooks, device=tokens.device)
191
+ )
192
+ audio_embeds = self.audio_embeddings(audio_tokens.view(-1)).reshape(
193
+ tokens.size(0), tokens.size(1), self.args.audio_num_codebooks, -1
194
+ )
195
+
196
+ return torch.cat([audio_embeds, text_embeds], dim=-2)
prompts/conversational_a.wav ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:356648c1bc6c1da7883004557e9b21a2ef7d01682d8b9d02d6dcb950b348b04f
3
+ size 2646044
prompts/conversational_b.wav ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c247153011385d33aaeed193adfec562c32182e2facd30cc8cd0b3e820e94afb
3
+ size 2646044
prompts/read_speech_a.wav ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:59480708f84c77ab2967d14d821c2ccade9d7761685d060575121f49a149005b
3
+ size 831412
prompts/read_speech_b.wav ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f582640265864499cbe6a8c687ea0f9e08e7fa41eeb2caa923d0a3bada55fcef
3
+ size 576052
prompts/read_speech_c.wav ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7da15ab3ee7f8bbc8abfce73ce65936a80a535ae4a86db2d9c4756caba69e9c3
3
+ size 385964
prompts/read_speech_d.wav ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:09cad0494f9d0038b0f0eb039f47d752c45e56d92679f96587e20f67b2c1b7d8
3
+ size 435884
requirements.txt ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ torch==2.4.0
2
+ torchaudio==2.4.0
3
+ tokenizers==0.21.0
4
+ transformers==4.49.0
5
+ huggingface_hub==0.28.1
6
+ spaces==0.32.0
7
+ gradio==5.20.1
8
+ moshi==0.2.2
9
+ torchtune==0.4.0
10
+ torchao==0.9.0
11
+ silentcipher @ git+https://github.com/SesameAILabs/silentcipher@master
watermarking.py ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import argparse
3
+
4
+ import silentcipher
5
+ import torch
6
+ import torchaudio
7
+
8
+ CSM_1B_HF_WATERMARK = list(map(int, os.getenv("WATERMARK_KEY").split(" ")))
9
+
10
+
11
+ def cli_check_audio() -> None:
12
+ parser = argparse.ArgumentParser()
13
+ parser.add_argument("--audio_path", type=str, required=True)
14
+ args = parser.parse_args()
15
+
16
+ check_audio_from_file(args.audio_path)
17
+
18
+
19
+ def load_watermarker(device: str = "cuda") -> silentcipher.server.Model:
20
+ model = silentcipher.get_model(
21
+ model_type="44.1k",
22
+ device=device,
23
+ )
24
+ return model
25
+
26
+
27
+ @torch.inference_mode()
28
+ def watermark(
29
+ watermarker: silentcipher.server.Model,
30
+ audio_array: torch.Tensor,
31
+ sample_rate: int,
32
+ watermark_key: list[int],
33
+ ) -> tuple[torch.Tensor, int]:
34
+ audio_array_44khz = torchaudio.functional.resample(audio_array, orig_freq=sample_rate, new_freq=44100)
35
+ encoded, _ = watermarker.encode_wav(audio_array_44khz, 44100, watermark_key, calc_sdr=False, message_sdr=36)
36
+
37
+ output_sample_rate = min(44100, sample_rate)
38
+ encoded = torchaudio.functional.resample(encoded, orig_freq=44100, new_freq=output_sample_rate)
39
+ return encoded, output_sample_rate
40
+
41
+
42
+ @torch.inference_mode()
43
+ def verify(
44
+ watermarker: silentcipher.server.Model,
45
+ watermarked_audio: torch.Tensor,
46
+ sample_rate: int,
47
+ watermark_key: list[int],
48
+ ) -> bool:
49
+ watermarked_audio_44khz = torchaudio.functional.resample(watermarked_audio, orig_freq=sample_rate, new_freq=44100)
50
+ result = watermarker.decode_wav(watermarked_audio_44khz, 44100, phase_shift_decoding=True)
51
+
52
+ is_watermarked = result["status"]
53
+ if is_watermarked:
54
+ is_csm_watermarked = result["messages"][0] == watermark_key
55
+ else:
56
+ is_csm_watermarked = False
57
+
58
+ return is_watermarked and is_csm_watermarked
59
+
60
+
61
+ def check_audio_from_file(audio_path: str) -> None:
62
+ watermarker = load_watermarker(device="cuda")
63
+
64
+ audio_array, sample_rate = load_audio(audio_path)
65
+ is_watermarked = verify(watermarker, audio_array, sample_rate, CSM_1B_HF_WATERMARK)
66
+
67
+ outcome = "Watermarked" if is_watermarked else "Not watermarked"
68
+ print(f"{outcome}: {audio_path}")
69
+
70
+
71
+ def load_audio(audio_path: str) -> tuple[torch.Tensor, int]:
72
+ audio_array, sample_rate = torchaudio.load(audio_path)
73
+ audio_array = audio_array.mean(dim=0)
74
+ return audio_array, int(sample_rate)
75
+
76
+
77
+ if __name__ == "__main__":
78
+ cli_check_audio()