Getting Started With The Bernie Mac Voice
If you are looking into Reginald Ballar The Bernie Mac Voice Of The Bernie Mac Shows Heart, you are probably trying to replicate the late comedian's vocal delivery for a project. This is a niche but surprisingly well-documented area of voice work. The approach generally involves either recording against a licensed performance or running your text through a trained voice model, depending on your budget and intended use. Most people come into this thinking it is a single file you download and run. It is not. What you are actually working with is a combination of source material, a voice model or licensed performer, and the post-processing required to make it sound consistent across a full script. I spent about three weeks last year working through this exact workflow for a parody project, so I will walk through what actually works and where people usually waste money. Let me explain the method first because the definition keeps getting mixed up online. You do not just paste text into a voice engine and get a finished product. The core process breaks down into four stages: sourcing clean reference audio, training or licensing the voice model, generating raw output, and then editing that output to remove artifacts and match pacing. I have seen people skip step one entirely and wonder why the result sounds like a robot reading a menu. The reference audio quality determines everything downstream. If your source clips have background noise, reverb from a studio audience, or compression artifacts baked in, the model will learn those flaws and reproduce them constantly.
I recommend starting with isolated vocal tracks if you can find them. Clean a cappella segments from The Bernie Mac Show episodes or standup recordings where there is no music bed underneath. One reliable trick is to run the raw video through a vocal isolation tool like those based on UVR5 or Demucs, then pick the cleanest 3 to 5 minute segments. You want variation in tone and pitch here, not just one monotone delivery. I used about 4 minutes of segmented audio for my own model run, and the output quality held up reasonably well for non-commercial use. Now, the counter-intuitive part that most tutorials miss. Shorter training runs often produce better results than long ones for this particular voice type. Bernie Mac's delivery has a very distinct cadence with sharp punctuation and frequent pauses. When you dump 30 minutes of raw audio into a trainer, the model tends to smooth over those abrupt shifts and create something that sounds more generic. I trained a baseline model on 30 minutes once and it sounded flat. I retrained with 4 minutes of carefully selected clips and the result was sharper and more recognizable. It took longer to curate the clips, but the final output was genuinely usable. There is a specific edge case I ran into that I want to mention because it will probably trip you up. When generating lines that contain heavy Southern inflection or rapid-fire punchlines, the model starts drifting into phonetic errors around the third or fourth sentence in a row. The energy drops, the vowels flatten out, and suddenly "ain't" sounds like "ant." I spent two days troubleshooting this before I realized the issue was sample rate mismatch between my source clips and the training config. My reference audio was at 22050 Hz but the trainer default was 16000 Hz. Once I matched both and also added a light noise floor around -40 dB to the training samples, the drift stopped. It sounds like a tiny detail but it makes the difference between a usable take and one you have to discard.
If you are not building a model from scratch and instead want to license a performance directly, Reginald Ballar's work is available through standard voice actor platforms and some independent booking sites. Rates vary, but you are typically looking at a flat fee for a session rather than a per-word rate. The advantage here is you get human performance with actual comedic timing baked in. The disadvantage is scheduling and the fact that some clients refuse to work with impersonations for trademark reasons. Check your intended distribution channel before committing funds. For the DIY route using open-source tools, RVC v2 remains the most accessible option for most people. It requires a GPU with at least 8 GB of VRAM for reasonable training times. You can run it on a CPU but expect several hours for a model that would take 20 minutes on a 3090. The interface itself is straightforward, but the preprocessing step is where most people give up. You need to clean your audio, remove silence gaps, and convert to WAV format before the trainer will accept it. A quick batch script using ffmpeg can handle the conversion, and deleting silent sections can be done with a tool like Audacity or a Python script using librosa. Here are the practical settings I ended up using after testing multiple configurations:
Get the Full Details
- Training epochs: 20 to 30. Going beyond 30 starts overfitting and the voice becomes strained on longer passages.
- Learning rate: 0.00015. Higher rates cause the model to chase artifacts instead of learning the actual timbre.
- Batch size: 4 if you have 8 GB VRAM, 8 if you have 12 GB or more.
- Pitch extraction: RMVPE gives noticeably cleaner results than the older PCENSR variant, especially for spoken word content.
After generation, do not skip the post-processing step. Raw outputs from voice models almost always need a bit of EQ work. I run the generated audio through a high-pass filter at 80 Hz to remove low-end rumble, then a gentle shelf cut around 300 Hz to reduce the boxy quality these models tend to introduce. A light compressor with a 3 dB gain reduction brings the dynamics back into a broadcast-friendly range. This usually takes about 10 to 15 minutes per minute of final audio, so budget your time accordingly. I should be straightforward about the limitations here. Even with optimal settings, this approach will not perfectly capture Bernie Mac's live standup energy. The model can approximate the vocal texture and general cadence, but the comedic rhythm, the look-before-you-speak pauses, and the crowd reaction timing are all lost. If your project depends on those elements, you are better off commissioning a human voice actor who specializes in impressions and giving them a written brief. The results will be noticeably better and you avoid any licensing gray areas. Another thing worth noting is the legal risk. Using a deceased performer's voice without permission exists in a legal vacuum that is currently being tested in courts. The right of publicity laws vary by state, and federal legislation on this topic is still evolving. If you are doing this for personal amusement, you are likely fine. If you plan to monetize the output, publish it publicly, or use it in advertising, consult someone who actually practices entertainment law rather than relying on forum advice.
For the actual download and setup, the RVC project repository on GitHub is the main source. The model files for specific voice impersonations are sometimes shared in communityDiscord servers and voice cloning forums. I found a few usable versions, but the quality varied wildly between uploads. My recommendation is to test any downloaded model on a short sample first before committing to a full script run. A bad model will waste hours of your time and the audio will not sound any better no matter how much you process it afterward. The whole process, from clean audio gathering to final mixed output, typically takes between 6 and 12 hours for a first-timer. After you have done it twice, you can compress that to about 3 hours. The learning curve is steeper on the technical side than the creative side, which is unusual for most voice work. You do not need to be an audio engineer, but you do need to be comfortable following technical instructions and troubleshooting when something goes wrong. If you end up sticking with this, keep a log of your training parameters and the reference clips you used. I went back and adjusted three settings on my second run based on notes I kept from the first, and the improvement was immediate. That habit alone will save you more time than any single tool or setting ever will.
