Dynamic Accent Pro v1.1 (experimental prompt-controlled Nigerian TTS); Flash-DPO default, Pro via --instruct
582aab9 verified | """Natural-language instruction templates for Flash-Instruct. | |
| labels: dict(lang=<ibo|yor|hau|pcm|eng|cs:<l1>+<l2>>, rate=slow|normal|fast|None, pause=few|normal|many|None, | |
| energy=calm|neutral|lively|None, gender=male|female|None) | |
| make(labels, rng) -> one instruction string (many surface forms per label combination). | |
| DEFAULT = the original CosyVoice3 system prompt; ~18% of training items keep it. | |
| """ | |
| import random | |
| DEFAULT = "You are a helpful assistant." | |
| LN = {"ibo": "Igbo", "yor": "Yoruba", "hau": "Hausa", "pcm": "Nigerian Pidgin", "eng": "Nigerian English"} | |
| RATE = { | |
| "slow": ["slowly", "at an unhurried pace", "at a slow, relaxed pace", "taking your time", "at a gentle, slow pace", | |
| "a little slower than usual", "slowly and deliberately"], | |
| "normal": ["at a natural pace", "at a normal speaking pace", "at a steady, everyday pace", "at a comfortable pace"], | |
| "fast": ["quickly", "at a brisk pace", "fast, like quick everyday conversation", "at a quick pace", | |
| "a little faster than usual", "briskly"], | |
| } | |
| PAUSE = { | |
| "few": ["with few pauses", "in one smooth flow with hardly any pauses", "without stopping much", | |
| "with minimal pausing", "running the phrases together"], | |
| "normal": ["with natural pauses", "with normal phrasing", "pausing where it feels natural"], | |
| "many": ["with plenty of pauses", "pausing often between phrases", "with frequent pauses", | |
| "with lots of breaks between phrases", "with deliberate pauses"], | |
| } | |
| ENERGY = { | |
| "calm": ["calm", "soft and calm", "gentle, even-toned", "relaxed and low-key", "quiet and steady"], | |
| "neutral": ["neutral", "plain, matter-of-fact", "even", "straightforward"], | |
| "lively": ["lively", "animated", "expressive and energetic", "bright and lively", "upbeat, expressive"], | |
| } | |
| GENDER = {"male": ["male", "man's"], "female": ["female", "woman's"]} | |
| def art(w): | |
| return ("an " if w[:1].lower() in "aeiou" else "a ") + w | |
| def lang_phrase(lang, rng): | |
| if lang.startswith("cs:"): | |
| a, b = lang[3:].split("+") | |
| A, B = LN.get(a, a), LN.get(b, b).replace("Nigerian English", "English") | |
| return rng.choice([f"an {A}-{B} mix" if A[0] in "AEIOU" else f"a {A}-{B} mix", f"mixed {A} and {B}", | |
| f"{A} with some {B} mixed in", f"code-switched {A} and {B}"]) | |
| return LN.get(lang, lang) | |
| def make(lab, rng=random): | |
| parts = {} | |
| if lab.get("rate"): | |
| parts["rate"] = rng.choice(RATE[lab["rate"]]) | |
| if lab.get("pause"): | |
| parts["pause"] = rng.choice(PAUSE[lab["pause"]]) | |
| if lab.get("energy"): | |
| parts["energy"] = rng.choice(ENERGY[lab["energy"]]) | |
| lang = lang_phrase(lab["lang"], rng) if lab.get("lang") else None | |
| g = rng.choice(GENDER[lab["gender"]]) if lab.get("gender") else None | |
| tone = parts.get("energy") | |
| adv = [parts[k] for k in ("rate", "pause") if k in parts] | |
| advs = ", ".join(adv) | |
| T = [] | |
| # template family 1: imperative sentence | |
| s = "Speak" | |
| if lang: | |
| s += f" in {lang}" + (f", sounding {tone}," if tone else "") | |
| elif tone: | |
| s += f" in {art(tone)} tone" | |
| if advs: | |
| s += f" {advs}" | |
| T.append(s + ".") | |
| # family 2: read-this style | |
| s = "Read this" | |
| if advs: | |
| s += f" {advs}" | |
| if tone: | |
| s += f", in {art(tone)} voice" | |
| if lang: | |
| s += f". The language is {lang}" | |
| T.append(s + ".") | |
| # family 3: descriptive/caption style | |
| d = [] | |
| if g: | |
| d.append(f"A {g} voice") | |
| else: | |
| d.append("A speaker") | |
| if lang: | |
| d.append(f"speaking {lang}") | |
| s = " ".join(d) | |
| if advs: | |
| s += f", {advs}" | |
| if tone: | |
| s += f"; the delivery is {tone}" if (lang or advs) else f" with {art(tone)} delivery" | |
| T.append(s + ".") | |
| # family 4: tag list | |
| tags = [] | |
| if lang: tags.append(f"language: {lang}") | |
| if "rate" in parts: tags.append(f"pace: {lab['rate']}") | |
| if "pause" in parts: tags.append(f"pauses: {lab['pause']}") | |
| if tone: tags.append(f"tone: {lab['energy']}") | |
| if g: tags.append(f"voice: {lab['gender']}") | |
| if tags: | |
| T.append("Style - " + "; ".join(tags) + ".") | |
| # family 5: polite request | |
| s = "Please say this" | |
| if lang: | |
| s += f" in {lang}" | |
| if tone: | |
| s += f" with {art(tone)} delivery" | |
| if advs: | |
| s += f", {advs}" | |
| T.append(s + ".") | |
| out = rng.choice(T) | |
| return out[0].upper() + out[1:] | |
| def sample_labels(full, rng, p=None): | |
| """drop each control independently so instructions mention subsets; keep at least one.""" | |
| p = p or dict(lang=0.55, rate=0.75, pause=0.7, energy=0.6, gender=0.3) | |
| while True: | |
| lab = {k: (v if (v and rng.random() < p.get(k, 0.5)) else None) for k, v in full.items()} | |
| if any(lab.values()): | |
| return lab | |
| if __name__ == "__main__": | |
| r = random.Random(0) | |
| full = dict(lang="cs:ibo+eng", rate="slow", pause="many", energy="calm", gender="female") | |
| for _ in range(12): | |
| print(make(sample_labels(full, r), r)) | |
| for l in ["ibo", "yor", "hau", "pcm", "eng"]: | |
| print(make(dict(lang=l, rate="fast", pause="few", energy="lively"), r)) | |