TallyReel blog

The best open-source text-to-speech for Hindi and Urdu: what we use and why

support@tallyreel.com (Danish) — Sun, 18 Oct 2026 12:00:00 GMT

VoxCPM2 was the best open model we could run locally. On 10 Hindi lines scored by Whisper it got 2.3 percent of letters wrong, against 4.1 percent for Kokoro and 5.1 percent for Veena, and it was about twice as fast as Veena on an M1 Pro. It also clones and designs voices. For Urdu it reads best from a Devanagari spelling: 6 percent of words wrong, against 28 percent when it read Urdu script directly.

What it really costs to render a faceless video on a cloud GPU

support@tallyreel.com (Danish) — Sun, 18 Oct 2026 12:00:00 GMT

A warm 30 second story Short with AI pictures and a natural voice took about 3 minutes on an L4 and about $0.04 of GPU time. The real bill came to about 1.5 times the GPU-only estimate once CPU, memory and idle time were counted, and with a 30 percent allowance for redos, the script AI, moderation and storage, the all-in cost is about $0.08 per Short. Long videos cost more than their length suggests.

Running FLUX image models locally on a Mac for video pictures

support@tallyreel.com (Danish) — Sun, 18 Oct 2026 12:00:00 GMT

FLUX.2 klein 4B clearly beat SDXL in all three art styles we tried, mainly because it follows long prompts: SDXL's 77-token limit cut off the character descriptions, so a character never appeared. In the shipped 4-bit MLX build, klein draws a scene in about 50 seconds on an M1 Pro and a whole five-scene video took 254 seconds. It is slower than SDXL's 30 seconds a scene, and worth it.

LTX vs Wan for image-to-video: why we parked moving scenes

support@tallyreel.com (Danish) — Sun, 18 Oct 2026 12:00:00 GMT

Not yet worth it for everyday stories. Moving two scenes of a six-scene story with Wan 2.2 TI2V-5B took the render from 3.5 to 8.6 minutes on an L4 GPU and the cost from about $0.08 to about $0.19, 2.4 times as much, and the result was not clearly better. LTX-2.3 was faster and sharper in our later test but merged two characters together, and FastWan took 7 minutes for a 2 second clip on a Mac. Stills stay the default.

We tested open-source lip-sync: MuseTalk vs EchoMimic for talking AI hosts

support@tallyreel.com (Danish) — Sun, 18 Oct 2026 12:00:00 GMT

MuseTalk 1.5 won clearly. On the same line and the same face, its mouth looked more natural, and it was far faster: a 43 second two-host conversation rendered in about 7.5 minutes on an M1 Pro, while EchoMimic V3 Flash took 14 minutes for 3.7 seconds of video. EchoMimic does move the head, but the overall result was worse, so TallyReel keeps MuseTalk for its hosts.

How we write a story in one AI call (and why five passes lost)

support@tallyreel.com (Danish) — Sun, 18 Oct 2026 12:00:00 GMT

One well-built call beat five to seven separate passes. On the same story ideas, the one-call writer scored about 7.3 out of 10 against about 5.0 for the multi-pass writer. Later, a version that lets the model think before writing won 7 of 8 blind comparisons, for about $0.001 more per story. The biggest gains came from a better single prompt and thinking time, not from more passes.