YarnGPT-local

Table of Contents

1. Model Summary
2. Model Description
3. Bias, Risks, and Limitations
   • Recommendations
4. Speech Samples
5. Training
6. Future Improvements
7. Citation
8. Credits & References

Model Summary

YarnGPT-local is a text-to-speech (TTS) model designed to synthesize Yoruba, Igbo and Hausa leveraging pure language modelling without external adapters or complex architectures, offering high-quality, natural, and culturally relevant speech synthesis for diverse applications.

How to use (on Google Colab)

The model can generate audio on its own but its better to use a voice to prompt the model, there are about 10 voices supported by default:

• hausa_female1
• hausa_female2
• hausa_male1
• hausa_male2
• igbo_female1
• igbo_female2
• igbo_male2
• yoruba_female1
• yoruba_female2
• yoruba_male2

Prompt YarnGPT-local

# clone the YarnGPT repo to get access to the `audiotokenizer`
!git clone https://github.com/saheedniyi02/yarngpt.git


# install some necessary libraries
!pip install outetts==0.2.3 uroman

#import some important packages 
import os
import re
import json
import torch
import inflect
import random
import uroman as ur
import numpy as np
import torchaudio
import IPython
from transformers import AutoModelForCausalLM, AutoTokenizer
from outetts.wav_tokenizer.decoder import WavTokenizer
from yarngpt.audiotokenizer import AudioTokenizerForLocal


# download the wavtokenizer weights and config (to encode and decode the audio)
!wget https://huggingface.co/novateur/WavTokenizer-medium-speech-75token/resolve/main/wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml
!gdown 1-ASeEkrn4HY49yZWHTASgfGFNXdVnLTt

# model path and wavtokenizer weight path (the paths are assumed based on Google colab, a different environment might save the weights to a different location).
hf_path="saheedniyi/YarnGPT-local"
wav_tokenizer_config_path="/content/wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml"
wav_tokenizer_model_path = "/content/wavtokenizer_large_speech_320_24k.ckpt"

# create the AudioTokenizer object 
audio_tokenizer=AudioTokenizerForLocal(
    hf_path,wav_tokenizer_model_path,wav_tokenizer_config_path
)

#load the model weights

model = AutoModelForCausalLM.from_pretrained(hf_path,torch_dtype="auto").to(audio_tokenizer.device)

# your input text
text="Ẹ maa rii pe lati bi ọsẹ meloo kan ni ijiroro ti wa lati ọdọ awọn ileeṣẹ wọnyi wi pe wọn fẹẹ ṣafikun si owo ipe pẹlu ida ọgọrun-un."

# creating a prompt, when creating a prompt, there is an optional `speaker_name` parameter
prompt=audio_tokenizer.create_prompt(text,"yoruba","yoruba_male2")

# tokenize the prompt
input_ids=audio_tokenizer.tokenize_prompt(prompt)

# generate output from the model, you can tune the `.generate` parameters as you wish
output  = model.generate(
            input_ids=input_ids,
            temperature=0.1,
            repetition_penalty=1.1,
            num_beams=4,
            max_length=4000,
        )

# convert the output to "audio codes"
codes=audio_tokenizer.get_codes(output)

# converts the codes to audio 
audio=audio_tokenizer.get_audio(codes)

# play the audio
IPython.display.Audio(audio,rate=24000)

# save the audio 
torchaudio.save(f"audio.wav", audio, sample_rate=24000)

Simple News-Reader for Local languages

# clone the YarnGPT repo to get access to the `audiotokenizer`
!git clone https://github.com/saheedniyi02/yarngpt.git


# install some necessary libraries
!pip install outetts uroman trafilatura pydub


#import important packages
import os
import re
import json
import torch
import inflect
import random
import requests
import trafilatura
import inflect
import uroman as ur
import numpy as np
import torchaudio
import IPython
from pydub import AudioSegment
from pydub.effects import normalize
from transformers import AutoModelForCausalLM, AutoTokenizer
from outetts.wav_tokenizer.decoder import WavTokenizer
from yarngpt.audiotokenizer import AudioTokenizer,AudioTokenizerForLocal

# download the `WavTokenizer` files
!wget https://huggingface.co/novateur/WavTokenizer-medium-speech-75token/resolve/main/wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml
!gdown 1-ASeEkrn4HY49yZWHTASgfGFNXdVnLTt

tokenizer_path="saheedniyi/YarnGPT-local"
wav_tokenizer_config_path="/content/wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml"
wav_tokenizer_model_path = "/content/wavtokenizer_large_speech_320_24k.ckpt"


audio_tokenizer=AudioTokenizerForLocal(
    tokenizer_path,wav_tokenizer_model_path,wav_tokenizer_config_path
       )

model = AutoModelForCausalLM.from_pretrained(tokenizer_path,torch_dtype="auto").to(audio_tokenizer.device)

# Split text into chunks
def split_text_into_chunks(text, word_limit=25):
  sentences=[sentence.strip() for sentence in text.split('.') if sentence.strip()]
  chunks=[]
  for sentence in sentences:
    chunks.append(".")
    sentence_splitted=sentence.split(" ")
    num_words=len(sentence_splitted)
    start_index=0
    if num_words>word_limit:
      while start_index<num_words:
        end_index=min(num_words,start_index+word_limit)
        chunks.append(" ".join(sentence_splitted[start_index:start_index+word_limit]))
        start_index=end_index
    else:
      chunks.append(sentence)
  return chunks

# reduce the speed of the audio, results from the local languages are always fast
def speed_change(sound, speed=0.9):
    # Manually override the frame_rate. This tells the computer how many
    # samples to play per second
    sound_with_altered_frame_rate = sound._spawn(sound.raw_data, overrides={
         "frame_rate": int(sound.frame_rate * speed)
      })
     # convert the sound with altered frame rate to a standard frame rate
     # so that regular playback programs will work right. They often only
     # know how to play audio at standard frame rate (like 44.1k)
    return sound_with_altered_frame_rate.set_frame_rate(sound.frame_rate)


page=requests.get("https://alaroye.org/a-maa-too-fo-ipinle-ogun-mo-omo-egbe-okunkun-meje-lowo-ti-te-bayii-omolola/")
content=trafilatura.extract(page.text)
chunks=split_text_into_chunks(content)


all_codes=[]
for i,chunk in enumerate(chunks):
  print(i)
  print("\n")
  print(chunk)
  if chunk==".":
    #add silence for 0.5 seconds if we encounter a full stop
    all_codes.extend([453]*38)
  else:
    prompt=audio_tokenizer.create_prompt(chunk,lang="yoruba",speaker_name="yoruba_female2")
    input_ids=audio_tokenizer.tokenize_prompt(prompt)
    output  = model.generate(
            input_ids=input_ids,
            temperature=0.1,
            repetition_penalty=1.1,
            max_length=4000,
            num_beams=5,
        )
    codes=audio_tokenizer.get_codes(output)
    all_codes.extend(codes)


audio=audio_tokenizer.get_audio(all_codes)

#display the output
IPython.display.Audio(audio,rate=24000)

#save audio
torchaudio.save(f"news1.wav", audio, sample_rate=24000)

#convert file to an `AudioSegment` object for furher processing
audio_dub=AudioSegment.from_file("news1.wav")

# reduce audio speed: it reduces quality also
speed_change(audio_dub,0.9)

Model Description

• Developed by: Saheedniyi
• Model type: Text-to-Speech
• Language(s) (NLP): Igbo, Yoruba, Hausa--> Speech
• Finetuned from: HuggingFaceTB/SmolLM2-360M
• Repository: YarnGPT Github Repository
• Paper: IN PROGRESS.
• Demo: 1) Prompt YarnGPT-local notebook 2) Simple news reader: YarnGPT-local

Uses

Generate yoruba, igbo and hausa speech for experimental purposes.

Out-of-Scope Use

The model is not suitable for generating speech in languages other than Yoruba, Igbo and Hausa.

Bias, Risks, and Limitations

• The model may not capture the full diversity of Nigerian accents and could exhibit biases based on the training dataset.
• The audio generated by the model are sometimes very fast and might need some post-processing to be done.
• The model doesn't take 'intonations' into account which sometimes leads to mispronounce meant of some words.
• Model doesn't respond to some prompt

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases, and limitations of the model. Feedback and diverse training data contributions are encouraged.

Speech Samples

Listen to samples generated by YarnGPT:

Training

Data

Trained on open source dataset on Yoruba, Igbo and Hausa.

Preprocessing

Audio files were preprocessed and resampled to 24Khz and tokenized using wavtokenizer.

Training Hyperparameters

• Number of epochs: 5
• batch_size: 4
• Scheduler: linear schedule with warmup for 4 epochs, then linear decay to zero for the last epoch
• Optimizer: AdamW (betas=(0.9, 0.95),weight_decay=0.01)
• Learning rate: 1*10^-3

Hardware

• GPUs: 1 A100 (google colab: 30 hours)

Software

• Training Framework: Pytorch

Future Improvements?

• Scaling up model size and training data
• Wrap the model around an API endpoint
• Voice cloning.
• Potential expansion into speech-to-speech assistant models

Citation [optional]

BibTeX:

@misc{yarngpt2025,
  author = {Saheed Azeez},
  title = {YarnGPT: Nigerian-Accented English Text-to-Speech Model},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/SaheedAzeez/yarngpt}
}

APA:

Saheed Azeez. (2025). YarnGPT-local: Nigerian languages Text-to-Speech Model. Hugging Face. Available at: https://huggingface.co/saheedniyi/YarnGPT-local

Credits & References

• OuteAI/OuteTTS-0.2-500M
• WavTokenizer
• CTC Forced Alignment
• Voicera