Voice cloning: per-case model and resource comparison
Same eight scripts and same local reference clip, synthesized on an NVIDIA RTX 2050 with 4,096 MiB physical VRAM. The source reference file is linked above.
Model file size is disk usage, not VRAM usage. OmniVoice rows report sampled total GPU and native-process peaks. Chatterbox reports per-case PyTorch allocated/reserved peaks; nvidia-smi totals, utilization and power are available only for the final three FP16 cases because the first-phase sample file was overwritten. Chatterbox changed from FP32 to FP16 + autocast after an out-of-memory error.
Timing boundaries differ: OmniVoice wall time includes the HTTP request and fresh process/model loading; Chatterbox generation time excludes its one-time download/load and includes reference conditioning. Treat timing as setup-specific, not a direct speed ranking. GPU energy is an estimate from sampled board power, not a wall-meter reading.
User's overall listening verdict: Chatterbox was not good overall; OmniVoice did better in several respects. Per-case quality ratings were not supplied.
Hindi-English code-switching
84 characters · seed 101 · OmniVoice language setting: Auto · Chatterbox language ID: hi
आज मैंने एक नया AI model try किया। Honestly, इसकी आवाज़ बिल्कुल मेरी जैसी लग रही है।
Listen for: Listen for accent changes, pronunciation of English words inside Hindi, skipped words, and preservation of speaker identity.
Serveurperso/OmniVoice-GGUF · Q8_0
Base model: k2-fsa/OmniVoice
Model files on disk
901.5 MiB total (626.0 MiB base + 275.5 MiB tokenizer)
Generation / output
9.14 s / 10.76 s · RTF 0.85
Peak GPU VRAM used
1,596 MiB total · 1,504 MiB OmniVoice process
Mean GPU utilization
73.0%
Peak board power / estimated GPU energy
43.6 W / 0.091 Wh
ResembleAI/chatterbox · Multilingual V3
Model files on disk
3,060.3 MiB total (T3 2,044.7 MiB + S3Gen 1,008.2 MiB + support 7.4 MiB)
Generation / output
18.33 s / 6.04 s · RTF 3.03
Peak PyTorch VRAM
3,284 MiB allocated · 3,450 MiB reserved · FP32
Peak total GPU VRAM
Unavailable: FP32-phase external GPU samples were not retained
GPU telemetry
Unavailable mean utilization · Unavailable peak power · Unavailable sampled board energy · 0 samples
Numbers, dates, currency and abbreviations
83 characters · seed 102 · OmniVoice language setting: en · Chatterbox language ID: en
My appointment is on October 21st at 9:45 AM. The bill is ₹1,249.50, including GST.
Listen for: Check October twenty-first, nine forty-five A.M., the rupee amount including fifty paise, and G-S-T. Alternate natural readings are acceptable if the value stays correct.
Serveurperso/OmniVoice-GGUF · Q8_0
Base model: k2-fsa/OmniVoice
Model files on disk
901.5 MiB total (626.0 MiB base + 275.5 MiB tokenizer)
Generation / output
10.28 s / 14.44 s · RTF 0.71
Peak GPU VRAM used
1,628 MiB total · 1,524 MiB OmniVoice process
Mean GPU utilization
70.6%
Peak board power / estimated GPU energy
44.3 W / 0.107 Wh
ResembleAI/chatterbox · Multilingual V3
Model files on disk
3,060.3 MiB total (T3 2,044.7 MiB + S3Gen 1,008.2 MiB + support 7.4 MiB)
Generation / output
9.59 s / 8.52 s · RTF 1.13
Peak PyTorch VRAM
3,332 MiB allocated · 3,558 MiB reserved · FP32
Peak total GPU VRAM
Unavailable: FP32-phase external GPU samples were not retained
GPU telemetry
Unavailable mean utilization · Unavailable peak power · Unavailable sampled board energy · 0 samples
Questions and emotion
89 characters · seed 103 · OmniVoice language setting: en · Chatterbox language ID: en
You actually finished it? That’s incredible! Wait… did you delete the original recording?
Listen for: Listen for a questioning first phrase, excitement, a pause, and concern. Punctuation alone is being tested; no emotion tags or style instructions are supplied.
Serveurperso/OmniVoice-GGUF · Q8_0
Base model: k2-fsa/OmniVoice
Model files on disk
901.5 MiB total (626.0 MiB base + 275.5 MiB tokenizer)
Generation / output
8.60 s / 10.47 s · RTF 0.82
Peak GPU VRAM used
1,606 MiB total · 1,490 MiB OmniVoice process
Mean GPU utilization
68.6%
Peak board power / estimated GPU energy
45.2 W / 0.091 Wh
ResembleAI/chatterbox · Multilingual V3
Model files on disk
3,060.3 MiB total (T3 2,044.7 MiB + S3Gen 1,008.2 MiB + support 7.4 MiB)
Generation / output
5.03 s / 4.00 s · RTF 1.26
Peak PyTorch VRAM
3,247 MiB allocated · 3,516 MiB reserved · FP32
Peak total GPU VRAM
Unavailable: FP32-phase external GPU samples were not retained
GPU telemetry
Unavailable mean utilization · Unavailable peak power · Unavailable sampled board energy · 0 samples
Names and technical vocabulary
79 characters · seed 104 · OmniVoice language setting: en · Chatterbox language ID: en
Jivraj is testing PostgreSQL, Kubernetes, and ElevenLabs on an NVIDIA RTX 2050.
Listen for: Check your name, PostgreSQL, Kubernetes, ElevenLabs, NVIDIA, R-T-X and the model number. Note spelling-out, omissions or odd stress.
Serveurperso/OmniVoice-GGUF · Q8_0
Base model: k2-fsa/OmniVoice
Model files on disk
901.5 MiB total (626.0 MiB base + 275.5 MiB tokenizer)
Generation / output
8.48 s / 10.92 s · RTF 0.78
Peak GPU VRAM used
1,592 MiB total · 1,500 MiB OmniVoice process
Mean GPU utilization
65.8%
Peak board power / estimated GPU energy
45.9 W / 0.090 Wh
ResembleAI/chatterbox · Multilingual V3
Model files on disk
3,060.3 MiB total (T3 2,044.7 MiB + S3Gen 1,008.2 MiB + support 7.4 MiB)
Generation / output
8.58 s / 7.76 s · RTF 1.11
Peak PyTorch VRAM
3,316 MiB allocated · 3,516 MiB reserved · FP32
Peak total GPU VRAM
Unavailable: FP32-phase external GPU samples were not retained
GPU telemetry
Unavailable mean utilization · Unavailable peak power · Unavailable sampled board energy · 0 samples
Consistency take 1
293 characters · seed 201 · OmniVoice language setting: en · Chatterbox language ID: en
I started this experiment with a simple question: can a small local model sound like me? Now I am checking whether it stays consistent. Listen to my accent, the pauses between sentences, and the way I finish each phrase. A familiar voice should remain familiar from the first word to the last.
Listen for: Compare takes 1–3 for identity, accent, pacing, completeness and noise. Different explicit seeds test variation across sampled takes.
Serveurperso/OmniVoice-GGUF · Q8_0
Base model: k2-fsa/OmniVoice
Model files on disk
901.5 MiB total (626.0 MiB base + 275.5 MiB tokenizer)
Generation / output
25.62 s / 34.93 s · RTF 0.73
Peak GPU VRAM used
1,618 MiB total · 1,514 MiB OmniVoice process
Mean GPU utilization
70.2%
Peak board power / estimated GPU energy
46.1 W / 0.289 Wh
ResembleAI/chatterbox · Multilingual V3
Model files on disk
3,060.3 MiB total (T3 2,044.7 MiB + S3Gen 1,008.2 MiB + support 7.4 MiB)
Generation / output
17.04 s / 14.52 s · RTF 1.17
Peak PyTorch VRAM
3,478 MiB allocated · 3,556 MiB reserved · FP32
Peak total GPU VRAM
Unavailable: FP32-phase external GPU samples were not retained
GPU telemetry
Unavailable mean utilization · Unavailable peak power · Unavailable sampled board energy · 0 samples
Consistency take 2
293 characters · seed 202 · OmniVoice language setting: en · Chatterbox language ID: en
I started this experiment with a simple question: can a small local model sound like me? Now I am checking whether it stays consistent. Listen to my accent, the pauses between sentences, and the way I finish each phrase. A familiar voice should remain familiar from the first word to the last.
Listen for: Compare against the identical script in takes 1 and 3; seed is the only requested generation setting changed.
Serveurperso/OmniVoice-GGUF · Q8_0
Base model: k2-fsa/OmniVoice
Model files on disk
901.5 MiB total (626.0 MiB base + 275.5 MiB tokenizer)
Generation / output
26.16 s / 35.40 s · RTF 0.74
Peak GPU VRAM used
1,618 MiB total · 1,514 MiB OmniVoice process
Mean GPU utilization
71.8%
Peak board power / estimated GPU energy
47.0 W / 0.297 Wh
ResembleAI/chatterbox · Multilingual V3
Model files on disk
3,060.3 MiB total (T3 2,044.7 MiB + S3Gen 1,008.2 MiB + support 7.4 MiB)
76.8% mean utilization · 53.9 W peak power · 0.119 Wh sampled board energy · 46 samples
Consistency take 3
293 characters · seed 203 · OmniVoice language setting: en · Chatterbox language ID: en
I started this experiment with a simple question: can a small local model sound like me? Now I am checking whether it stays consistent. Listen to my accent, the pauses between sentences, and the way I finish each phrase. A familiar voice should remain familiar from the first word to the last.
Listen for: Compare against takes 1 and 2. Different WAV hashes alone do not establish an audible difference.
Serveurperso/OmniVoice-GGUF · Q8_0
Base model: k2-fsa/OmniVoice
Model files on disk
901.5 MiB total (626.0 MiB base + 275.5 MiB tokenizer)
Generation / output
26.14 s / 35.28 s · RTF 0.74
Peak GPU VRAM used
1,618 MiB total · 1,514 MiB OmniVoice process
Mean GPU utilization
68.3%
Peak board power / estimated GPU energy
47.3 W / 0.302 Wh
ResembleAI/chatterbox · Multilingual V3
Model files on disk
3,060.3 MiB total (T3 2,044.7 MiB + S3Gen 1,008.2 MiB + support 7.4 MiB)
94.5% mean utilization · 53.8 W peak power · 0.101 Wh sampled board energy · 34 samples
Long narration
1553 characters · seed 301 · OmniVoice language setting: en · Chatterbox language ID: en
This morning, I decided to take the long way to the neighbourhood market. Usually, I rush through the same streets with headphones on, thinking about the next task. Today, I left the headphones at home and paid attention to the sounds around me.
A bicycle bell rang behind me. Someone was watering plants on a balcony. At a small tea stall, two friends were arguing about whether the best ideas arrive early in the morning or late at night. I wanted to join their discussion, but the smell of fresh bread pulled me towards the bakery instead.
Inside, the shopkeeper asked, “The usual?” I almost said yes. Then I noticed a handwritten sign beside a tray of cinnamon rolls. “Actually, I'll try something different today,” I replied. It was a small decision, but it changed the rhythm of the morning.
On the walk home, I thought about the project waiting on my laptop. Yesterday, a short recording had become a voice that sounded surprisingly familiar: my own. The exciting part was hearing it speak a sentence I had never recorded. The harder question was whether it could keep sounding like me through a whole story.
So this paragraph is part of the experiment. Does the voice stay steady? Do the questions sound like questions? Are the pauses comfortable, or do they break the flow? Listen closely as one paragraph becomes the next.
By the time I reached home, the bread was still warm. I made a cup of tea, opened the laptop, and wrote down one simple rule: a convincing first sentence is a beginning. A convincing final sentence is the real test.
Listen for: Listen for voice drift, paragraph joins, missing or repeated sentences, uneven pacing, unnatural pauses, and whether the final paragraph still sounds like you.
Serveurperso/OmniVoice-GGUF · Q8_0
Base model: k2-fsa/OmniVoice
Model files on disk
901.5 MiB total (626.0 MiB base + 275.5 MiB tokenizer)
Generation / output
134.31 s / 186.06 s · RTF 0.72
Peak GPU VRAM used
1,632 MiB total · 1,528 MiB OmniVoice process
Mean GPU utilization
71.1%
Peak board power / estimated GPU energy
49.9 W / 1.582 Wh
ResembleAI/chatterbox · Multilingual V3
Model files on disk
3,060.3 MiB total (T3 2,044.7 MiB + S3Gen 1,008.2 MiB + support 7.4 MiB)