MosesJoshuaCoker

Novax-TTS-v3

Model Details

Model name: Novax-Krio-TTS-v3
Base model: canopylabs/orpheus-3b-0.1-ft (Llama-3.2-3B-Instruct)
Architecture: Transformer (Llama), 3B parameters
Fine-tuning method: QLoRA (4-bit, LoRA rank 64)
Audio codec: SNAC 24 kHz (7 tokens/frame)
Output sample rate: 24 kHz, mono, PCM int16
Language: Krio (Sierra Leone / Liberia Creole English)
Adapter size: ~150 MB

Quick Start

Installation

bash

pip install unsloth snac soundfile torch transformers

Usage

Load the adapter on top of the base Orpheus model:

python

import torchfrom snac import SNACfrom unsloth import FastLanguageModelimport soundfile as sf # Load model + adaptermodel, tokenizer = FastLanguageModel.from_pretrained(    model_name="MosesJoshuaCoker/Novax-Krio-TTS-v3",    max_seq_length=4096,    dtype=None,    load_in_4bit=True,)FastLanguageModel.for_inference(model) # Load SNAC decodersnac = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval().cuda() # Constants (fixed by Orpheus)END_OF_TEXT, START_OF_SPEECH, END_OF_SPEECH = 128009, 128257, 128258START_OF_HUMAN, END_OF_HUMAN, START_OF_AI = 128259, 128260, 128261AUDIO_CODE_OFFSET, CODEBOOK = 128266, 4096 def synthesize(text, temperature=0.6, top_p=0.95, max_tokens=2400):    """Generate Krio speech from text."""    ids = tokenizer(text, return_tensors="pt").input_ids    start = torch.tensor([[START_OF_HUMAN]], dtype=torch.long)    end = torch.tensor([[END_OF_TEXT, END_OF_HUMAN, START_OF_AI, START_OF_SPEECH]],                       dtype=torch.long)    input_ids = torch.cat([start, ids, end], dim=1).cuda()        gen = model.generate(        input_ids=input_ids,        attention_mask=torch.ones_like(input_ids),        max_new_tokens=max_tokens,        do_sample=True,        temperature=temperature,        top_p=top_p,        eos_token_id=END_OF_SPEECH,        pad_token_id=tokenizer.pad_token_id,    )        seq = gen[0]    sos = (seq == START_OF_SPEECH).nonzero()    seq = seq[sos[-1].item() + 1:] if len(sos) else seq    raw = [t.item() - AUDIO_CODE_OFFSET for t in seq if t.item() >= AUDIO_CODE_OFFSET]    raw = raw[: (len(raw) // 7) * 7]        cap = lambda v: max(0, min(CODEBOOK - 1, v))    l1, l2, l3 = [], [], []    for i in range(len(raw) // 7):        l1.append(cap(raw[7*i]))        l2.append(cap(raw[7*i+1] - 1*CODEBOOK))        l3.append(cap(raw[7*i+2] - 2*CODEBOOK))        l3.append(cap(raw[7*i+3] - 3*CODEBOOK))        l2.append(cap(raw[7*i+4] - 4*CODEBOOK))        l3.append(cap(raw[7*i+5] - 5*CODEBOOK))        l3.append(cap(raw[7*i+6] - 6*CODEBOOK))        layers = [torch.tensor(l1).unsqueeze(0).cuda(),              torch.tensor(l2).unsqueeze(0).cuda(),              torch.tensor(l3).unsqueeze(0).cuda()]        with torch.inference_mode():        wav = snac.decode(layers).squeeze().cpu().numpy()    return wav # Exampletext = "kushɛ, aw di bɔdi? wi de tray fɔ mek krio tɛkst-tu-spich voice."wav = synthesize(text)sf.write("krio_sample.wav", wav, 24000)

Or use the standalone CLI:

bash

python inference.py --adapter MosesJoshuaCoker/Novax-Krio-TTS-v3 \    --text "kushɛ, aw di bɔdi?" --out krio.wav

Training Data

Dataset: MosesJoshuaCoker/30_hours_krio_voice
Total audio: ~30 hours
Clips: 6,807 examples (average ~16 seconds each)
Language: Krio (orthography with diacritics: ɛ, ɔ, ŋ, etc.)
Splits: single train split
Audio format: originally 16 kHz mono PCM, resampled to 24 kHz for SNAC

Training Details

Table with columns: Parameter, Value
Parameter	Value
Base model	`canopylabs/orpheus-3b-0.1-ft`
LoRA rank	64
LoRA alpha	64
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Quantization	4-bit (NF4)
Optimizer	AdamW8bit
Learning rate	2e-4 (cosine decay)

Audio Tokenization

Each audio frame is encoded into 7 SNAC tokens across 3 resolution levels, with per-position offsets:

text

Position 0: base_offset + 0 * CODEBOOKPosition 1: base_offset + 1 * CODEBOOKPosition 2: base_offset + 2 * CODEBOOKPosition 3: base_offset + 3 * CODEBOOKPosition 4: base_offset + 4 * CODEBOOKPosition 5: base_offset + 5 * CODEBOOKPosition 6: base_offset + 6 * CODEBOOK

where base_offset = 128266 and CODEBOOK = 4096. This maps to the Orpheus vocabulary which already contains these audio codes (no embedding resize needed).

Limitations & Known Issues

Audio quality: SNAC is lossy; some high-frequency detail is lost compared to the original 30 h dataset.
Sequence length: max 4096 tokens (~26 seconds of audio). Longer clips require a higher MAX_SEQ_LEN setting (and more GPU VRAM).
Single speaker: trained on a single-speaker dataset. Voice characteristics are consistent but limited to that speaker's prosody and accent.
Inference speed: ~0.034 steps/sec on a T4, so a 10-second clip takes ~1.5 minutes to generate (not real-time).
Repetition: occasionally generates repetitive or looping speech. Lower temperature (0.4–0.5) and raise repetition_penalty (1.2–1.3) to mitigate.
Orthography: only trained on Krio text. Input text should use Krio orthography (ɛ, ɔ, ŋ, etc.). English text will be read as English.

Intended Use

Primary: synthesizing fluent Krio speech from written text for accessibility, education, or media production.
Suitable for: offline TTS applications, research, language preservation, assistive technology.
Not suitable for: real-time applications (inference is slow), high-fidelity audio production (use a vocoder post-processor if needed).

Ethical Considerations

Single-speaker bias: all training data is from a single speaker. Outputs will sound like that speaker.
Accent & dialect: Krio has regional and sociolinguistic variation; this model captures one variety.
Misuse: do not use to impersonate a specific individual or create deepfake speech. The model does not preserve speaker identity from the dataset beyond its inherent training characteristics.

How to Cite

bibtex

@model{novax-krio-tts-v3,  author = {Moses Joshua Coker},  title = {Novax Krio TTS v3: A LoRA-Finetuned Orpheus-3B for Krio Speech Synthesis},  url = {https://huggingface.co/MosesJoshuaCoker/Novax-Krio-TTS-v3},  year = {2026}}

References

Orpheus TTS: canopyai/Orpheus-TTS on GitHub.
SNAC: Siuzdak et al., "SNAC: Unified Neural Audio Codec", 2024.
Krio language: Wikipedia.
Unsloth: Fast QLoRA finetuning — unslothai/unsloth.

License

MIT. See LICENSE in the repo.

Questions or feedback? Open an issue on the model repo.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

MosesJoshuaCoker

Model Tree

Base

unsloth/orpheus-3b-0.1-ft

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

Model Details

Model name: Novax-Krio-TTS-v3
Base model: canopylabs/orpheus-3b-0.1-ft (Llama-3.2-3B-Instruct)
Architecture: Transformer (Llama), 3B parameters
Fine-tuning method: QLoRA (4-bit, LoRA rank 64)
Audio codec: SNAC 24 kHz (7 tokens/frame)
Output sample rate: 24 kHz, mono, PCM int16
Language: Krio (Sierra Leone / Liberia Creole English)
Adapter size: ~150 MB

Quick Start

Installation

bash

pip install unsloth snac soundfile torch transformers

Usage

Load the adapter on top of the base Orpheus model:

python

import torchfrom snac import SNACfrom unsloth import FastLanguageModelimport soundfile as sf # Load model + adaptermodel, tokenizer = FastLanguageModel.from_pretrained(    model_name="MosesJoshuaCoker/Novax-Krio-TTS-v3",    max_seq_length=4096,    dtype=None,    load_in_4bit=True,)FastLanguageModel.for_inference(model) # Load SNAC decodersnac = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").eval().cuda() # Constants (fixed by Orpheus)END_OF_TEXT, START_OF_SPEECH, END_OF_SPEECH = 128009, 128257, 128258START_OF_HUMAN, END_OF_HUMAN, START_OF_AI = 128259, 128260, 128261AUDIO_CODE_OFFSET, CODEBOOK = 128266, 4096 def synthesize(text, temperature=0.6, top_p=0.95, max_tokens=2400):    """Generate Krio speech from text."""    ids = tokenizer(text, return_tensors="pt").input_ids    start = torch.tensor([[START_OF_HUMAN]], dtype=torch.long)    end = torch.tensor([[END_OF_TEXT, END_OF_HUMAN, START_OF_AI, START_OF_SPEECH]],                       dtype=torch.long)    input_ids = torch.cat([start, ids, end], dim=1).cuda()        gen = model.generate(        input_ids=input_ids,        attention_mask=torch.ones_like(input_ids),        max_new_tokens=max_tokens,        do_sample=True,        temperature=temperature,        top_p=top_p,        eos_token_id=END_OF_SPEECH,        pad_token_id=tokenizer.pad_token_id,    )        seq = gen[0]    sos = (seq == START_OF_SPEECH).nonzero()    seq = seq[sos[-1].item() + 1:] if len(sos) else seq    raw = [t.item() - AUDIO_CODE_OFFSET for t in seq if t.item() >= AUDIO_CODE_OFFSET]    raw = raw[: (len(raw) // 7) * 7]        cap = lambda v: max(0, min(CODEBOOK - 1, v))    l1, l2, l3 = [], [], []    for i in range(len(raw) // 7):        l1.append(cap(raw[7*i]))        l2.append(cap(raw[7*i+1] - 1*CODEBOOK))        l3.append(cap(raw[7*i+2] - 2*CODEBOOK))        l3.append(cap(raw[7*i+3] - 3*CODEBOOK))        l2.append(cap(raw[7*i+4] - 4*CODEBOOK))        l3.append(cap(raw[7*i+5] - 5*CODEBOOK))        l3.append(cap(raw[7*i+6] - 6*CODEBOOK))        layers = [torch.tensor(l1).unsqueeze(0).cuda(),              torch.tensor(l2).unsqueeze(0).cuda(),              torch.tensor(l3).unsqueeze(0).cuda()]        with torch.inference_mode():        wav = snac.decode(layers).squeeze().cpu().numpy()    return wav # Exampletext = "kushɛ, aw di bɔdi? wi de tray fɔ mek krio tɛkst-tu-spich voice."wav = synthesize(text)sf.write("krio_sample.wav", wav, 24000)

Or use the standalone CLI:

bash

python inference.py --adapter MosesJoshuaCoker/Novax-Krio-TTS-v3 \    --text "kushɛ, aw di bɔdi?" --out krio.wav

Training Data

Dataset: MosesJoshuaCoker/30_hours_krio_voice
Total audio: ~30 hours
Clips: 6,807 examples (average ~16 seconds each)
Language: Krio (orthography with diacritics: ɛ, ɔ, ŋ, etc.)
Splits: single train split
Audio format: originally 16 kHz mono PCM, resampled to 24 kHz for SNAC

Training Details

Table with columns: Parameter, Value
Parameter	Value
Base model	`canopylabs/orpheus-3b-0.1-ft`
LoRA rank	64
LoRA alpha	64
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Quantization	4-bit (NF4)
Optimizer	AdamW8bit
Learning rate	2e-4 (cosine decay)

Audio Tokenization

Each audio frame is encoded into 7 SNAC tokens across 3 resolution levels, with per-position offsets:

text

Position 0: base_offset + 0 * CODEBOOKPosition 1: base_offset + 1 * CODEBOOKPosition 2: base_offset + 2 * CODEBOOKPosition 3: base_offset + 3 * CODEBOOKPosition 4: base_offset + 4 * CODEBOOKPosition 5: base_offset + 5 * CODEBOOKPosition 6: base_offset + 6 * CODEBOOK

Position 0: base_offset + 0 * CODEBOOKPosition 1: base_offset + 1 * CODEBOOKPosition 2: base_offset + 2 * CODEBOOKPosition 3: base_offset + 3 * CODEBOOKPosition 4: base_offset + 4 * CODEBOOKPosition 5: base_offset + 5 * CODEBOOKPosition 6: base_offset + 6 * CODEBOOK

where base_offset = 128266 and CODEBOOK = 4096. This maps to the Orpheus vocabulary which already contains these audio codes (no embedding resize needed).

Limitations & Known Issues

Audio quality: SNAC is lossy; some high-frequency detail is lost compared to the original 30 h dataset.
Sequence length: max 4096 tokens (~26 seconds of audio). Longer clips require a higher MAX_SEQ_LEN setting (and more GPU VRAM).
Single speaker: trained on a single-speaker dataset. Voice characteristics are consistent but limited to that speaker's prosody and accent.
Inference speed: ~0.034 steps/sec on a T4, so a 10-second clip takes ~1.5 minutes to generate (not real-time).
Repetition: occasionally generates repetitive or looping speech. Lower temperature (0.4–0.5) and raise repetition_penalty (1.2–1.3) to mitigate.
Orthography: only trained on Krio text. Input text should use Krio orthography (ɛ, ɔ, ŋ, etc.). English text will be read as English.

Intended Use

Primary: synthesizing fluent Krio speech from written text for accessibility, education, or media production.
Suitable for: offline TTS applications, research, language preservation, assistive technology.
Not suitable for: real-time applications (inference is slow), high-fidelity audio production (use a vocoder post-processor if needed).

Ethical Considerations

Single-speaker bias: all training data is from a single speaker. Outputs will sound like that speaker.
Accent & dialect: Krio has regional and sociolinguistic variation; this model captures one variety.
Misuse: do not use to impersonate a specific individual or create deepfake speech. The model does not preserve speaker identity from the dataset beyond its inherent training characteristics.

How to Cite

bibtex

@model{novax-krio-tts-v3,  author = {Moses Joshua Coker},  title = {Novax Krio TTS v3: A LoRA-Finetuned Orpheus-3B for Krio Speech Synthesis},  url = {https://huggingface.co/MosesJoshuaCoker/Novax-Krio-TTS-v3},  year = {2026}}

References

Orpheus TTS: canopyai/Orpheus-TTS on GitHub.
SNAC: Siuzdak et al., "SNAC: Unified Neural Audio Codec", 2024.
Krio language: Wikipedia.
Unsloth: Fast QLoRA finetuning — unslothai/unsloth.

License

MIT. See LICENSE in the repo.

Questions or feedback? Open an issue on the model repo.

Novax-TTS-v3

README

Model Details

Quick Start

Installation

Usage

Training Data

Training Details

Audio Tokenization

Limitations & Known Issues

Intended Use

Ethical Considerations

How to Cite

References

License

Explore FriendliAI today

README

Model Details

Quick Start

Installation

Usage

Training Data

Training Details

Audio Tokenization

Limitations & Known Issues

Intended Use

Ethical Considerations

How to Cite

References

License