Ming-3.0-tts-0.5B vs today's voices

Brown = what the sites use today (ElevenLabs / OpenAI / in-clip voices). Green = Ming running on our own GPU (INCREDIBLE, RTX 5060 Ti), $0 per line. Same text and same character description on both sides. Not wired into any site yet.

22Ming lines
171 sof audio
166 sto make it
1.03xreal time overall
3.1 GBVRAM
44.1 kHzoutput

1. Finished film narration: film 1827, "Kael the Jester and the baker who forgot her laughter"

Today: StoryCast narrator voice George (ElevenLabs JBFqnCBsd6RMkjVDRZzb, eleven_v4_turbo), described in config as "British man, warm, captivating storyteller". Ming designs a voice from that description; no reference audio.

Line 1: "Dawn light spills over cobblestones as steam rises from Hilda's oven chimney, curling like forgotten sighs."
Ming: same 6-word description
7.8 s audiomade in 7.1 s1.10x real time
Ming: description written out in full
6.7 s audiomade in 6.1 s1.10x real time
Line 6: "Then, pop, the dough landed on her nose. She blinked. She sneezed. And for the first time in years, Hilda the Baker laughed..."
Ming: full description
13.1 s audiomade in 17.0 s0.77x real time

2. Two characters in a supporting-cast film (film 1829, Kael + Hilda)

Today every StoryCast line uses the film's single narrator voice, so Kael and Hilda are both read by George. Ming gives each character a voice designed from their look: Kael = child jester, Hilda = late-60s gruff baker.

Kael: "I've tried juggling, puns, even a squirrel in a tiny hat, but Hilda's laugh's gone colder than yesterday's bread."
Ming: Kael (designed child voice)
6.5 s audiomade in 5.8 s1.12x real time
Hilda (new line; there is no Hilda voice today): "Out of my kitchen, jester, before I bake you into a pie. And mind the flour on your way out."
Ming: Hilda (designed older woman)
5.5 s audiomade in 4.9 s1.12x real time

3. AIMusicVideo storyboard voice lines

Commercial speech scenes (Seedance 2.0) call ElevenLabs eleven_flash_v2_5 (voice picked by male/female hash, here Adam). None are kept on disk, so this one line was made fresh for the test. The editor explainer uses OpenAI tts-1 (nova / onyx); those are real saved files.

Commercial line: "Meet the sneaker that keeps up with your city. Light, bold, and ready when you are."
Ming: designed "confident young American man"
4.1 s audiomade in 3.7 s1.12x real time
Ming: built-in US_James
4.9 s audiomade in 4.5 s1.10x real time
Editor explainer (design job 92): "Change the entire city background where she is and place her on a beach setting..."
Ming: built-in US_Emma
6.6 s audiomade in 6.0 s1.09x real time
Editor explainer (design job 89): "Turn this into a rainy night with wet reflections on the floor."
Ming: designed "deep calm American man"
3.8 s audiomade in 3.6 s1.07x real time

4. Ad narrator lines (demos.jason.com UGC / commercials)

Today demos ads have no separate narrator track: LTX/H3 make up the voice inside the clip, and the avatar lane uses P-Video Avatar's own voices. The ad line below was made once with ElevenLabs (the same setup AIMV commercials use) for comparison. It ran to 21 s, not 15. The two UGC lines are real saved clips (audio pulled from the video).

Ad narrator: "Tired of cold coffee by 9 a.m.? Meet the Ember mug... save $4.50 on your first mug, with free shipping all through 2026."
Ming: designed "upbeat lifestyle narrator"
15.5 s audiomade in 14.3 s1.08x real time
spoken as: Tired of cold coffee by nine A M? Meet the Ember mug. It keeps your drink at the exact temperature you like for up to eighty minutes. Just set it, sip it, and forget the microwave. Order today and save four dollars and fifty cents on your first mug, with free shipping all through twenty twenty-six.
Ming: built-in US_Stella
16.5 s audiomade in 15.1 s1.09x real time
spoken as: Tired of cold coffee by nine A M? Meet the Ember mug. It keeps your drink at the exact temperature you like for up to eighty minutes. Just set it, sip it, and forget the microwave. Order today and save four dollars and fifty cents on your first mug, with free shipping all through twenty twenty-six.
UGC (design job 260, LTX in-clip voice): "Oh my god, I can't believe this actually works!"
Ming: designed "excited young woman on her phone"
3.6 s audiomade in 3.3 s1.09x real time
UGC (design job 223, P-Video Avatar Zephyr): "Okay, real talk. I bought this on a whim and it is my favorite thing now."
Ming: designed "casual young woman"
4.3 s audiomade in 3.9 s1.11x real time

5. DomainMagic voice-over

Promo films have music only today. The DomainMagic builder's site "voice intro" uses ElevenLabs Rachel eleven_turbo_v2_5 (real file: agent-lab.test). The promo line is written from the bostonplumbers.com promo film's own on-screen copy.

Site voice intro: "Welcome to Arcade.tv. Play the best classic arcade games online. Press Start and relive the golden age of gaming."
Ming: designed "friendly website welcome"
6.7 s audiomade in 6.0 s1.12x real time
Promo film voice-over (new; promos are silent today): "Dripping faucet? Time to find help. Find the right plumbing help in Boston..."
Ming: designed "local-business promo announcer"
9.5 s audiomade in 9.0 s1.06x real time

6. Voice clone from a 10-second clip of our own narration

Reference: film 1827 line 5, a 10.0 s George narration we generated ("So Kael waited, until noon..."). Ming clones that voice and reads new text.

Reference clip (10.0 s)
Line 1 again: "Dawn light spills over cobblestones..."
Ming: clone of the 10 s clip
6.4 s audiomade in 5.9 s1.08x real time
Kael's line: "I've tried juggling, puns, even a squirrel in a tiny hat..."
Ming: clone of the 10 s clip
6.7 s audiomade in 6.0 s1.12x real time

7. Inline tags: laugh, sigh, hesitation, breath

Tags go right in the text: [Laughter] [Sigh] [Uhm] [Breathing] [Surprise-oh] [Dissatisfaction-hnn]... ElevenLabs flash/turbo have no such tags.

George clone: "[Laughter] Oh, you should have seen her face! [Sigh] But the oven had already gone cold. [Uhm] And then, well... [Breathing] the dough landed right on her nose."
Ming: George clone + tags
9.7 s audiomade in 8.8 s1.10x real time
Hilda: "[Dissatisfaction-hnn] Jesters. [Sigh] Always jesters. [Laughter] Fine, fine, one scone."
Ming: Hilda designed voice + tags
5.3 s audiomade in 10.5 s0.50x real time
first use also designs the voice on a plain line (included in the time)
Ming: cloned from Hilda's first line + tags
6.6 s audiomade in 6.0 s1.10x real time
Built-in voice: "[Laughter] You actually believed that? [Sigh] Okay. [Uhm] Let me start over. [Surprise-oh] Oh! The map was never wrong."
Ming: US_Stella + tags
8.3 s audiomade in 7.4 s1.12x real time

8. Numbers and abbreviations

The model's own number cleanup (pynini) won't install on Windows, so the service spells out money, years, times, percents, ordinals, phone numbers and abbreviations itself first.

"Doors open at 7:30 pm on Oct. 31st. Tickets are $12.99, or 2 for $20, and 15% goes to charity. Call 555-0134."
Ming: built-in US_James
12.7 s audiomade in 11.6 s1.10x real time
spoken as: Doors open at seven thirty P M on October thirty-first. Tickets are twelve dollars and ninety-nine cents, or two for twenty dollars and fifteen percent goes to charity. Call five five five, zero one three four.

Only two paid lines were made for this page (sections 3 and 4, ElevenLabs flash, 340 characters). Every other "today" clip was already on disk.