SpeakSpace-Assistant-v1-3B

Alpha AI (www.alphaai.biz) fine-tuned canopylabs/orpheus-3b-0.1-ft to create SpeakSpace-Assistant-v1-3B — an English-only, single-speaker voice assistant model. The fine-tune uses custom voice recordings plus the Elise dataset (~3 hours, single-speaker English speech). Transcripts were augmented with emotion/expression tags like <sigh> and <laughs>, added as special tokens in the Orpheus tokenizer.

⚠️ Important: This model is intended for research, prototyping, and internal product demos. Do not use it to impersonate a real person without explicit consent. Review base-model and dataset licenses before commercial use.

---

TL;DR

• Base: canopylabs/orpheus-3b-0.1-ft (~3B params).
• Data: Custom Alpha AI dataset + MrDragonFox/Elise (English, ~3 hours).
• Objective: Produce natural, expressive speech with inline emotion cues (<laughs>, <sigh>).
• Language: English only.
• Repo: Suggested as alpha-ai/SpeakSpace-Assistant-v1-3B.

---

Intended Use & Limitations

Intended use:

• Internal voice assistants and demos.
• Research on expressive TTS and emotion-tag-conditioned speech.
• Applications where transcripts include small expressive markers.

Limitations:

• Not multi-speaker or multilingual.
• Quality limited by dataset size (~3 hrs + custom data).
• Requires Orpheus vocoder/decoder to convert tokens to waveform.
• Do not deploy for impersonation without explicit consent.

---

Model Details

• Family: Orpheus 3B (decoder-based speech model).
• Tokenizer: Extended with special tokens (<laughs>, <sigh>).
• Fine-tuning: Supervised finetuning on audio–transcript pairs.
• Output: Discrete audio tokens; decode with Orpheus vocoder.

---

Data

Sources:

• Alpha AI custom speech dataset.
• MrDragonFox/Elise (~3 hrs English single-speaker).

Preprocessing:

• Aligned utterances with transcripts.
• Expression tags inserted inline.
• Special tokens added to tokenizer.

---

Prompt & Input Format

Model accepts text input with optional inline expressions:

Hello! <laughs> I can help with your schedule today.

Workflow: tokenize → generate audio tokens → decode via vocoder.

---

Training Summary

• Objective: Predict audio tokens from transcripts (with expression markers).
• Loss: Causal LM loss.
• Optimizer: AdamW or AdamW-8bit (please add exact values).
• Hyperparameters: Learning rate, batch size, gradient accumulation, seed — to be filled with actual values.

---

Evaluation

Recommended:

• MOS (Mean Opinion Score): naturalness & expressiveness.
• Speaker similarity: ABX or MOS vs. ground truth.
• Intelligibility: WER via ASR.
• Emotion accuracy: Human rating of <laughs>, <sigh> cues.

Add quantitative results when available.

---

Safety & Responsible Use

• Use only with documented consent for training voices.
• Guard against impersonation risks.
• Consider watermarking or metadata tagging for provenance.
• Do not generalize beyond training speaker’s identity.

---

License & Attribution

• Base model: canopylabs/orpheus-3b-0.1-ft (review base license).
• Dataset: MrDragonFox/Elise (check dataset license).
• Fine-tune: Ensure compatibility of licenses.

Suggested citation:

SpeakSpace-Assistant-v1-3B — fine-tune of canopylabs/orpheus-3b-0.1-ft on Alpha AI custom dataset + MrDragonFox/Elise.

---

Acknowledgements

• canopylabs — Orpheus base model.
• MrDragonFox — Elise dataset.
• Alpha AI research & engineering team.

---

Contact

Questions, issues, or collaborations:

• Open a discussion on the Hugging Face repo.
• Enterprise contact (Alpha AI): www.alphaai.biz | corporate@alphaai.biz
• Enterprise contact (SpeakSpace): www.speakspace.co | connect@speakspace.co