GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
MosesJoshuaCoker
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
MosesJoshuaCoker
Model Tree
canopylabs/orpheus-3b-0.1-ft
Input Modalities
Output Modalities
Supported Functionality
bash
pip install "transformers>=4.43" snac torch soundfilepython
import torch, soundfile as sffrom snac import SNACfrom transformers import AutoModelForCausalLM, AutoTokenizer REPO = "MosesJoshuaCoker/Novax_Krio_V3"device = "cuda" if torch.cuda.is_available() else "cpu" tokenizer = AutoTokenizer.from_pretrained(REPO)model = AutoModelForCausalLM.from_pretrained(REPO, torch_dtype=torch.float16).to(device).eval()snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval().to(device) # Orpheus special tokens (fixed) + the voice name this model was trained with.SOH, EOT, EOH, SOA, SOS = 128259, 128009, 128260, 128261, 128257EOS_SPEECH, PAD_ID, AUDIO_BASE = 128258, 128263, 128266VOICE = "krio" @torch.inference_mode()def decode_frames(aud): """SNAC audio-token ids (already offset by AUDIO_BASE) -> 24 kHz waveform.""" l1, l2, l3 = [], [], [] for i in range(0, len(aud) - len(aud) % 7, 7): f = aud[i:i + 7] l1.append(f[0]) l2 += [f[1] - 4096, f[4] - 4 * 4096] l3 += [f[2] - 2 * 4096, f[3] - 3 * 4096, f[5] - 5 * 4096, f[6] - 6 * 4096] codes = [torch.tensor(c, device=device).unsqueeze(0).clamp(0, 4095) for c in (l1, l2, l3)] return snac_model.decode(codes)[0, 0].float().cpu().numpy() @torch.inference_mode()def say(text, max_seconds=20, temperature=0.6, top_p=0.95, repetition_penalty=1.3): prompt = torch.tensor([[SOH] + tokenizer.encode(f"{VOICE}: {text}") + [EOT, EOH, SOA, SOS]], device=device) out = model.generate( input_ids=prompt, attention_mask=torch.ones_like(prompt), max_new_tokens=int(max_seconds) * 84, # ~84 audio tokens per second do_sample=True, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty, # keep >= 1.1 or Orpheus loops eos_token_id=EOS_SPEECH, pad_token_id=PAD_ID, )[0].tolist() audio_ids = [t - AUDIO_BASE for t in out[prompt.shape[1]:] if t >= AUDIO_BASE] return decode_frames(audio_ids) wav = say("Kushe! Aw di bodi?")sf.write("krio.wav", wav, 24000)print("wrote krio.wav")temperature 0.4-0.7 and repetition_penalty 1.1-1.4. Raise the penalty if speech loops or babbles;
lower it if prosody sounds clipped or robotic.import torch, soundfile as sffrom snac import SNACfrom transformers import AutoModelForCausalLM, AutoTokenizer REPO = "MosesJoshuaCoker/Novax_Krio_V3"device = "cuda" if torch.cuda.is_available() else "cpu" tokenizer = AutoTokenizer.from_pretrained(REPO)model = AutoModelForCausalLM.from_pretrained(REPO, torch_dtype=torch.float16).to(device).eval()snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval().to(device) # Orpheus special tokens (fixed) + the voice name this model was trained with.SOH, EOT, EOH, SOA, SOS = 128259, 128009, 128260, 128261, 128257EOS_SPEECH, PAD_ID, AUDIO_BASE = 128258, 128263, 128266VOICE = "krio" @torch.inference_mode()def decode_frames(aud): """SNAC audio-token ids (already offset by AUDIO_BASE) -> 24 kHz waveform.""" l1, l2, l3 = [], [], [] for i in range(0, len(aud) - len(aud) % 7, 7): f = aud[i:i + 7] l1.append(f[0]) l2 += [f[1] - 4096, f[4] - 4 * 4096] l3 += [f[2] - 2 * 4096, f[3] - 3 * 4096, f[5] - 5 * 4096, f[6] - 6 * 4096] codes = [torch.tensor(c, device=device).unsqueeze(0).clamp(0, 4095) for c in (l1, l2, l3)] return snac_model.decode(codes)[0, 0].float().cpu().numpy() @torch.inference_mode()def say(text, max_seconds=20, temperature=0.6, top_p=0.95, repetition_penalty=1.3): prompt = torch.tensor([[SOH] + tokenizer.encode(f"{VOICE}: {text}") + [EOT, EOH, SOA, SOS]], device=device) out = model.generate( input_ids=prompt, attention_mask=torch.ones_like(prompt), max_new_tokens=int(max_seconds) * 84, # ~84 audio tokens per second do_sample=True, temperature=temperature, top_p=top_p, repetition_penalty=repetition_penalty, # keep >= 1.1 or Orpheus loops eos_token_id=EOS_SPEECH, pad_token_id=PAD_ID, )[0].tolist() audio_ids = [t - AUDIO_BASE for t in out[prompt.shape[1]:] if t >= AUDIO_BASE] return decode_frames(audio_ids) wav = say("Kushe! Aw di bodi?")sf.write("krio.wav", wav, 24000)print("wrote krio.wav")