A 16-bit quantized and merged, memory-efficient fine-tuned version of canopylabs/orpheus-3b-0.1-ft, optimized with Unsloth and LoRA, for expressive multilingual TTS, especially in low-resource African languages. This model provides the following capabilities:
This model was trained on a parallel text-speech dataset totaling over 300 hours (75k samples) of Nigerian-accented and low-resource language audio (Igbo, Yoruba, Hausa). A key part of the dataset comes from AfroVoices' transcription of real-world YouTube data (denoted as Random speaker, ~100+ hrs). To preserve and enhance multilingual capabilities while avoiding catastrophic forgetting, we included synthetic speech-text data sampled from the original 8 Orpheus voices using the default emotional prompts. The final training set also included new speakers like:
This model sets state-of-the-art performance on low-resource Multilingual TTS tasks across African languages (see training stats below)
The default Orpheus-TTS model released by Canopy Labs supports the below voices and emotions:
Voices: tara, leah, jess, leo, dan, mia, zac, and zoe.
Emotions: <laugh>, <chuckle>, <sigh>, <cough>, <sniffle>, <groan>, <yawn>, and <gasp>.
Through synthetic data generation and addition, our finetuned model also maintains these voices and emotions. Please visit the default model's card for mor info on voices and emotions.
🎧 Listen to samples generated by Hypa Orpheus-TTS
| Text Input | Audio Output | Language | Voice |
|---|---|---|---|
| I am cooking for guests tomorrow and need to know how to make aioli. Can you give me a step-by-step recipe. | English | Emmanuel | |
| Ina dafa abinci don bakin gobe kuma ina bukatar sanin yadda ake yin ailoli. Za ka iya ba ni girke-gireken matakan daya bayan daya? | Hausa | Emmanuel | |
| Ina dafa abinci don bakin gobe kuma ina bukatar sanin yadda ake yin ailoli. Za ka iya ba ni girke-gireken matakan daya bayan daya? | Hausa | Eniola | |
| Èmi máa se oúnjẹ fún àwọn àlejò l'ọ́la mo sì nílò láti mọ bí wọn ti ńṣe aioli. Ṣe o lè fún mi ni àwọn ìlànà ìdáná ẹlẹ́sẹẹsẹ? | Yoruba | Eniola | |
| I am cooking for guests tomorrow and need to know how to make aioli. Can you give me a step-by-step recipe. | English | Eniola | |
| M na-esi nri maka ndị ọbịa echi ma achọ ịmata otú esi esi aioli. Ị nwere ike inye m usoro ntụziaka? | Igbo | Eniola | |
| M na-esi nri maka ndị ọbịa echi ma achọ ịmata otú esi esi aioli. Ị nwere ike inye m usoro ntụziaka? | Igbo | Lovelyn | |
| I am cooking for guests tomorrow and need to know how to make aioli. Can you give me a step-by-step recipe. | English | Lovelyn |
| Step | Training Loss | Validation Loss |
|---|---|---|
| 5,000 | 3.9496 | 3.8790 |
| 10,000 | 3.8863 | 3.79497 |
| 15,000 | 3.8544 | 3.75323 |
We plan to open-source the full dataset shortly similar to the Hypa_Fleurs initiative.
This model is released under an Open Source License (apache-2.0). Please refer to the LICENSE file for full details.
When using this model in your work, please cite both this model as well as the base canopylabs/orpheus-3b-0.1-ft model as follows:
@misc{canopylabsorpheus,
title={Orpheus-3b-0.1-ft: A Multilingual Text-to-Speech Model},
author={Canopy Labs},
year={2025},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/canopylabs/orpheus-3b-0.1-ft}},
note={Fine-tuned version of Orpheus for expressive TTS}
}
@misc{hypaorpheus4bit,
title={Hypa_Orpheus-3b-0.1-ft (LoRA-4bit)},
author={Hypa AI},
year={2025},
note={Fine-tuned Orpheus TTS on African languages},
url={https://huggingface.co/hypaai/Hypa_Orpheus-3b-0.1-ft-unsloth-bnb-4bit}
}
For any questions, issues, or contributions, please open an issue in this repository or contact [email protected]. Contributions are welcome!
By making Hypa_Orpheus available, we hope to empower research and development in multilingual speech technologies for African languages.
Hypa AI remains steadfast in its mission to pioneer intelligent solutions that are not just technologically advanced but are also culturally aware, ensuring that the future of AI is as diverse and inclusive as the world it serves.
AfroVoices, a subsidiary of Hypa AI, is dedicated to amplifying African voices, languages, and cultures in the intelligence age. Focused on bridging the digital representation gap, AfroVoices curates datasets and resources for African languages, promoting inclusivity and cultural appreciation in AI technologies. Their mission goes beyond technological innovation, aiming to celebrate the richness of African linguistic diversity on a global stage.
Download the nedded packages.
%%capture
import os
if "COLAB_" not in "".join(os.environ.keys()):
!pip install unsloth
else:
# Do this only in Colab notebooks! Otherwise use pip install unsloth
!pip install --no-deps bitsandbytes accelerate xformers==0.0.29.post3 peft trl==0.15.2 triton cut_cross_entropy unsloth_zoo
!pip install sentencepiece protobuf datasets huggingface_hub hf_transfer
!pip install --no-deps unsloth
!pip install snac
Download the models (both the SNAC encoder/decoder as well as our finetuned Hypa_Orpheus).
import torch
from snac import SNAC
from unsloth import FastLanguageModel
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "hypaai/Hypa_Orpheus-3b-0.1-ft-unsloth-merged_16bit",
max_seq_length= 2048, # Choose any for long context!
dtype = dtype,
load_in_4bit = load_in_4bit,
#token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
snac_model = snac_model.to("cuda")
Create your text prompt, select the voice, and pass through the Model.
prompts = [
"""Mo nífẹ̀ẹ́sí láti ṣe Ph.D sùgbọ́n mi ò ì tíì pinnu ẹ̀ka tí màá ṣe. Àwọn anfaani tí óń dé oríṣiríṣi àwọn olùgbọ́ káàkiri àgbáyé wo ni mo ní""",
]
chosen_voice = "Eniola" # None for single-speaker
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
snac_model.to("cpu")# Moving snac_model cuda to cpu
prompts_ = [(f"{chosen_voice}: " + p) if chosen_voice else p for p in prompts]
all_input_ids = []
for prompt in prompts_:
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
all_input_ids.append(input_ids)
start_token = torch.tensor([[ 128259]], dtype=torch.int64) # Start of human
end_tokens = torch.tensor([[128009, 128260]], dtype=torch.int64) # End of text, End of human
all_modified_input_ids = []
for input_ids in all_input_ids:
modified_input_ids = torch.cat([start_token, input_ids, end_tokens], dim=1) # SOH SOT Text EOT EOH
all_modified_input_ids.append(modified_input_ids)
all_padded_tensors = []
all_attention_masks = []
max_length = max([modified_input_ids.shape[1] for modified_input_ids in all_modified_input_ids])
for modified_input_ids in all_modified_input_ids:
padding = max_length - modified_input_ids.shape[1]
padded_tensor = torch.cat([torch.full((1, padding), 128263, dtype=torch.int64), modified_input_ids], dim=1)
attention_mask = torch.cat([torch.zeros((1, padding), dtype=torch.int64), torch.ones((1, modified_input_ids.shape[1]), dtype=torch.int64)], dim=1)
all_padded_tensors.append(padded_tensor)
all_attention_masks.append(attention_mask)
all_padded_tensors = torch.cat(all_padded_tensors, dim=0)
all_attention_masks = torch.cat(all_attention_masks, dim=0)
input_ids = all_padded_tensors.to("cuda")
attention_mask = all_attention_masks.to("cuda")
generated_ids = model.generate(
input_ids=input_ids,
attention_mask=attention_mask,
max_new_tokens=1200,
do_sample=True,
temperature=0.6,
top_p=0.95,
repetition_penalty=1.1,
num_return_sequences=1,
eos_token_id=128258,
use_cache = True
)
token_to_find = 128257
token_to_remove = 128258
token_indices = (generated_ids == token_to_find).nonzero(as_tuple=True)
if len(token_indices[1]) > 0:
last_occurrence_idx = token_indices[1][-1].item()
cropped_tensor = generated_ids[:, last_occurrence_idx+1:]
else:
cropped_tensor = generated_ids
mask = cropped_tensor != token_to_remove
processed_rows = []
for row in cropped_tensor:
masked_row = row[row != token_to_remove]
processed_rows.append(masked_row)
code_lists = []
for row in processed_rows:
row_length = row.size(0)
new_length = (row_length // 7) * 7
trimmed_row = row[:new_length]
trimmed_row = [t - 128266 for t in trimmed_row]
code_lists.append(trimmed_row)
def redistribute_codes(code_list):
layer_1 = []
layer_2 = []
layer_3 = []
for i in range((len(code_list)+1)//7):
layer_1.append(code_list[7*i])
layer_2.append(code_list[7*i+1]-4096)
layer_3.append(code_list[7*i+2]-(2*4096))
layer_3.append(code_list[7*i+3]-(3*4096))
layer_2.append(code_list[7*i+4]-(4*4096))
layer_3.append(code_list[7*i+5]-(5*4096))
layer_3.append(code_list[7*i+6]-(6*4096))
codes = [torch.tensor(layer_1).unsqueeze(0),
torch.tensor(layer_2).unsqueeze(0),
torch.tensor(layer_3).unsqueeze(0)]
# codes = [c.to("cuda") for c in codes]
audio_hat = snac_model.decode(codes)
return audio_hat
my_samples = []
for code_list in code_lists:
samples = redistribute_codes(code_list)
my_samples.append(samples)
from IPython.display import display, Audio
if len(prompts) != len(my_samples):
raise Exception("Number of prompts and samples do not match")
else:
for i in range(len(my_samples)):
print(prompts[i])
samples = my_samples[i]
display(Audio(samples.detach().squeeze().to("cpu").numpy(), rate=24000))
# Clean up to save RAM
del my_samples,samples
Download the nedded packages.
%%capture
!pip install snac ipywebrtc
Download the Models (SNAC & Hypa_Orpheus)
import torch
from transformers import AutoModelForCausalLM, Trainer, TrainingArguments, AutoTokenizer
from snac import SNAC
# Loads the pre-trained SNAC model and moves it to the CPU.
snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
snac_model = snac_model #.to("cpu")
print("We have loaded the Encoder/Decoder model to the cpu, to use vram - use the gpu for faster inference")
# Loading the Orpheus Model and Tokenizer, moving the model to the GPU for faster inference
model_name = "hypaai/Hypa_Orpheus-3b-0.1-ft-unsloth-merged_16bit"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16)
model.cuda()
tokenizer = AutoTokenizer.from_pretrained(model_name)
Create Prompt(s) and Select Voice & Emotions as needed.
# List of supported voices in Orpheus-TTS
voices = [
"Eniola", "tara", # Female, conversational, clear
"Moyo", "leah", # Female, warm, gentle
"Gift", "jess", # Female, energetic, youthful
"Prince", "leo", # Male, authoritative, deep
"Emmanuel", "dan", # Male, friendly, casual
"Cynthia", "mia", # Female, professional, articulate
"Kolade", "zac", # Male, enthusiastic, dynamic
"Lovelyn", "zoe" # Female, calm, soothing
]
# List of supported emotion tags in Orpheus-TTS
emotions = [
"<laugh>", # Laughter
"<chuckle>", # Soft chuckle
"<sigh>", # Sighing
"<cough>", # Coughing
"<sniffle>", # Sniffling
"<groan>", # Groaning
"<yawn>", # Yawning
"<gasp>" # Gasping
]
# Creating Prompts
prompts = [
"Hey there my name is Eniola 9000, and I'm a speech generation model that can sound like a person.",
# "I've also been taught to understand and produce paralinguistic things like sighing, or chuckling, or yawning!",
# "I live in San Francisco, and have, uhm let's see, 3 billion 7 hundred ... well, lets just say a lot of parameters.",
]
chosen_voice = "Eniola" # "tara" # see github for other voices
prompts = [f"{chosen_voice}: " + p for p in prompts] # Creating the prompts (as a batch)
print(prompts)
Tokenize prompt(s) into inputIDs, pad, and create attention masks.
# Tokenizing each prompt into input IDs.
all_input_ids = []
for prompt in prompts:
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
all_input_ids.append(input_ids)
# Adds special tokens to mark the beginning and end of each prompt
start_token = torch.tensor([[128259]], dtype=torch.int64) # Start of human
end_tokens = torch.tensor([[128009, 128260]], dtype=torch.int64) # End of text, End of human
all_modified_input_ids = []
for input_ids in all_input_ids:
modified_input_ids = torch.cat([start_token, input_ids, end_tokens], dim=1) # SOH SOT Text EOT EOH
all_modified_input_ids.append(modified_input_ids)
# Padding All sequences to same length and creating corresponding attention masks
all_padded_tensors = []
all_attention_masks = []
max_length = max([modified_input_ids.shape[1] for modified_input_ids in all_modified_input_ids])
for modified_input_ids in all_modified_input_ids:
padding = max_length - modified_input_ids.shape[1]
# Left Padding
padded_tensor = torch.cat([torch.full((1, padding), 128263, dtype=torch.int64), modified_input_ids], dim=1)
attention_mask = torch.cat([torch.zeros((1, padding), dtype=torch.int64), torch.ones((1, modified_input_ids.shape[1]), dtype=torch.int64)], dim=1)
all_padded_tensors.append(padded_tensor)
all_attention_masks.append(attention_mask)
all_padded_tensors = torch.cat(all_padded_tensors, dim=0)
all_attention_masks = torch.cat(all_attention_masks, dim=0)
# Moving all padded sequences to GPU for Faster computation
input_ids = all_padded_tensors.to("cuda")
attention_mask = all_attention_masks.to("cuda")
Generate Output Tokens from the and Parse output tokens as speech
print("*** Model.generate is slow - see vllm implementation on github for realtime streaming and inference")
print("*** Increase/decrease inference params for more expressive less stable generations")
# Generating Output Tokens
with torch.no_grad():
generated_ids = model.generate(
input_ids=input_ids,
attention_mask=attention_mask,
max_new_tokens=1200,
do_sample=True,
temperature=0.6,
top_p=0.95,
repetition_penalty=1.1,
num_return_sequences=1,
eos_token_id=128258,
)
# Processing Generated Tokens (Parse Output as speech)
token_to_find = 128257 # Start of Audio token (relevant output)
token_to_remove = 128258 # End/ Terminal Token (End of Audio/ relevant output)
token_indices = (generated_ids == token_to_find).nonzero(as_tuple=True)
print(token_indices)
# Slices the tensor to exclude unwanted tokens.
if len(token_indices[1]) > 0:
last_occurrence_idx = token_indices[1][-1].item()
cropped_tensor = generated_ids[:, last_occurrence_idx+1:]
else:
cropped_tensor = generated_ids
# mask = cropped_tensor != token_to_remove
# Storing the cleaned-up token sequences#
processed_rows = []
for row in cropped_tensor:
masked_row = row[row != token_to_remove]
processed_rows.append(masked_row)
# Preparing (Audio Codes) the token sequences for audio decoding by trimming and adjusting token values.
code_lists = []
for row in processed_rows:
row_length = row.size(0) # Determines the length of the token sequence.
new_length = (row_length // 7) * 7 # Ensures the sequence length is a multiple of 7, as required by the decoder.
trimmed_row = row[:new_length]
trimmed_row = [t - 128266 for t in trimmed_row] # Adjusts token values to match the expected input range for the decoder.
code_lists.append(trimmed_row)
Decode Outputs with SNAC Decoder
# Processes the token sequences into the format expected by the SNAC decoder:
def redistribute_codes(code_list):
""" Reorganizes the flattened token list into three separate layers, adjusting each token's value to align with the decoder's expectations"""
layer_1 = [] # Coarsest layer
layer_2 = [] # Intermediate layer
layer_3 = [] # Finest layer
num_groups = (len(code_list) + 1) // 7 #Calculate the number of complete 7-token groups in the code_list
for i in range(num_groups):
idx = 7 * i # starting index for the current group
# Layer 1 receives the first token of the group
layer_1.append(code_list[idx])
# Layer 2 receives the second token, adjusted by subtracting 4096
layer_2.append(code_list[idx + 1] - 4096)
# Layer 3 receives the third and fourth tokens, adjusted by subtracting 8192 and 12288 respectively
layer_3.append(code_list[idx+2]-(2*4096))
layer_3.append(code_list[idx+3]-(3*4096))
# Layer 2 receives the fifth token, adjusted by subtracting 16384
layer_2.append(code_list[idx+4]-(4*4096))
# Layer 3 receives the sixth and seventh tokens, adjusted by subtracting 20480 and 24576 respectively
layer_3.append(code_list[idx+5]-(5*4096))
layer_3.append(code_list[idx+6]-(6*4096))
codes = [
torch.tensor(layer_1).unsqueeze(0), # Shape: (1, len(layer_1))
torch.tensor(layer_2).unsqueeze(0), # Shape: (1, len(layer_2))
torch.tensor(layer_3).unsqueeze(0) # Shape: (1, len(layer_3))
] # Convert the lists to PyTorch tensors and add a batch dimension
audio_hat = snac_model.decode(codes) # Decode the structured codes into an audio waveform using the SNAC model
return audio_hat
my_samples = []
for code_list in code_lists:
samples = redistribute_codes(code_list) # Generates audio samples from the processed token sequences
my_samples.append(samples)
# Display Audio
from IPython.display import display, Audio
if len(prompts) != len(my_samples):
raise Exception("Number of prompts and samples do not match")
else:
for i in range(len(my_samples)):
print(prompts[i])
samples = my_samples[i]
display(Audio(samples.detach().squeeze().to("cpu").numpy(), rate=24000))
This llama based model was trained 2x faster with Unsloth and Huggingface's TRL library.
Base model
meta-llama/Llama-3.2-3B-Instruct