Dogs Don T Tell Jokes

A Practical Walkthrough for Getting Real Results With Dogs Don T Tell Jokes

Dogs Don T Tell Jokes is a voice cloning and text-to-speech platform that lets you upload audio recordings and then generate speech in that same voice from any written text. It works by extracting a speaker embedding from your uploaded files and running it through a neural vocoder. The result can sound surprisingly close to the source, but it also has a long list of things that can go wrong if you are not paying attention. I started using it about a year ago when a client needed narrated content in a voice that sounded consistent across dozens of scripts. I signed up, went to the dashboard, and uploaded about six minutes of clean audio from the target speaker. The reference audio needs to be clear — no background music, no heavy reverb, no overlapping dialogue. I learned that the hard way. Once the upload finished processing, which took maybe three to five minutes for a short clip, I typed a script into the text box and hit generate. The first output came back in under thirty seconds on their servers. You get free credits when you sign up, and paid tiers scale from there depending on how much generation you need. The pricing is roughly in the range of a few cents per minute of output audio at the lower tiers.

Here is what most people skip, and it matters: after generating the audio, you should always listen back at the original pacing, not just the final mix. The TTS engine will sometimes drag out consonants or compress syllables in ways that sound fine at first pass but become obvious on a second listen. I found that adjusting the punctuation in my input text — adding commas where a natural pause would go, breaking long sentences into shorter ones — improved intelligibility more than any settings tweak.

The specific problem I ran into and the workaround

Early on, I uploaded a reference voice recorded in a home studio, and the generated output had a strange metallic undertone that was not present in the original recording. The voice sounded cloned correctly, but there was this synthetic shimmer on the sibilants. I spent about two hours going back and forth trying different enhancement filters inside the platform. Nothing worked. The issue turned out to be that the reference audio had a high-frequency noise floor from an inexpensive USB microphone. The model picked up that noise and baked it into the speaker embedding, then reproduced it in every generation. The workaround was straightforward: I ran the reference audio through a noise reduction pass in Audacity first, trimmed the noise gate tightly around the actual speech, and re-uploaded the cleaned version. The metallic artifact disappeared entirely. This took maybe ten minutes total and saved me from either paying for a professional voice actor or abandoning the project. If you ever run into similar artifacts, check your reference audio before you blame the platform. Run it through a spectral view in any DAW and look for consistent frequency bands that do not belong to the voice. That is usually your culprit.

Things the documentation does not emphasize enough

One thing nobody warns you about: the model does not handle strong regional accents consistently unless your reference audio contains one. If you upload a neutral-accent reference and ask it to read text with phonetic structures that belong to a different dialect, the output can sound flat or oddly flattened. I discovered this when a colleague asked me to generate Scottish-flavored narration from a reference I had recorded in London. The result sounded like someone reading Scottish text with a London accent. You need a reference that matches the phonetic profile of what you are asking it to say. Another detail: duration matters more than quality in some cases. A two-minute reference will often produce more stable output than a thirty-second one, even if the thirty seconds is studio-quality. The model needs enough phonetic variety in the embedding to generalize properly. Short clips tend to produce voices that sound correct on familiar words but distort on less common syllable combinations.

When this tool actually fails you

Dogs Don T Tell Jokes is not a solution for everything. It struggles with emotionally nuanced delivery. If you need a voice that sounds genuinely angry, whispering, or sobbing, the output tends to sound robotic no matter what you do. The model generates monotone emotional approximations at best. For commercial work requiring dramatic variation, you are better off hiring a voice actor or using a different platform that offers explicit emotion controls. It also does not handle code-switching well. If your script mixes languages mid-sentence, the output can stumble or default to one language throughout. I once had a bilingual Spanish-English script where the model just flattened everything into one accent pattern. Splitting the script by language and generating separately worked around this, but it adds time to the workflow. Long-form generation above about ten minutes of continuous output tends to show artifacts creeping in toward the end. The model loses consistency over extended passages. For longer content, break it into segments and stitch them together in a DAW afterward. You will save yourself a lot of re-generation time.

Where to find Dogs Don T Tell Jokes

You can access the platform directly through their official website. They offer a free tier with limited credits for testing, which is worth using before committing to a paid plan. I recommend generating a short test clip with your own voice or a public-domain reference before purchasing anything, just to confirm the output quality meets your needs. Their support response time is reasonable, and they do publish regular updates to the model, so performance does improve over time if you run into issues early on.