Impact Signals TTS Bake-off Soundboard

Commercial-safe candidates only. Built 2026-05-05T16:58:55. Kokoro clips are local same-text renders; the other clips are official model-author demos for style triage.
Models5
Playable clips13
Commercial5/5 OK
Default pickKokoro

Kokoro-82M

Apache-2.0Commercial OK5 clips
Role: Default production narrator / preset two-voice option
Evidence: Local render on Impact Signals sample text
Recommendation: Best default replacement now: fast, stable, permissive preset voices.
am_liam — Male default candidate
same Impact Signals text / local render · 16.45s
Impact Signals is tracking how artificial intelligence is moving from demonstration projects into public health, disaster response, and humanitarian operations. The replacement voice should sound calm, credible, and clear. It should pronounce organizations, numbers, and policy terms without skipping words.
local Kokoro-82M ONNX assets
am_michael — Male alternate
same Impact Signals text / local render · 20.63s
Impact Signals is tracking how artificial intelligence is moving from demonstration projects into public health, disaster response, and humanitarian operations. The replacement voice should sound calm, credible, and clear. It should pronounce organizations, numbers, and policy terms without skipping words.
local Kokoro-82M ONNX assets
af_bella — Female cohost candidate
same Impact Signals text / local render · 20.12s
Impact Signals is tracking how artificial intelligence is moving from demonstration projects into public health, disaster response, and humanitarian operations. The replacement voice should sound calm, credible, and clear. It should pronounce organizations, numbers, and policy terms without skipping words.
local Kokoro-82M ONNX assets
af_heart — Female high-grade candidate
same Impact Signals text / local render · 19.07s
Impact Signals is tracking how artificial intelligence is moving from demonstration projects into public health, disaster response, and humanitarian operations. The replacement voice should sound calm, credible, and clear. It should pronounce organizations, numbers, and policy terms without skipping words.
local Kokoro-82M ONNX assets
am_liam + af_bella — two preset voice dialogue
same Impact Signals dialogue / local render · 27.63s
Impact Signals is tracking how artificial intelligence is moving from demonstration projects into public health, disaster response, and humanitarian operations. / The replacement voice should sound calm, credible, and clear. It should pronounce organizations, numbers, and policy terms without skipping words. / For production, we care about commercial licensing, transcript fidelity, natural pacing, and whether the audio can pass automated guardrails every day.
local Kokoro-82M ONNX assets

VibeVoice

MITCommercial OK3 clips
Role: Long-form multi-speaker podcast candidate
Evidence: Official public demo samples downloaded from Microsoft GitHub Pages
Recommendation: Strong candidate to evaluate next with local inference; official demo suggests podcast fit.
2-person podcast: See You Again
official demo sample / not same Impact Signals text · 61.33s
Official demo text from model authors; use for voice/style only, not final transcript-fidelity certainty.
https://microsoft.github.io/VibeVoice/assets/audio/2p_see_u_again.mp3
2-person argument/dialogue
official demo sample / not same Impact Signals text · 68.53s
Official demo text from model authors; use for voice/style only, not final transcript-fidelity certainty.
https://microsoft.github.io/VibeVoice/assets/audio/2p_argument.mp3
3-person GPT-5 discussion
official demo sample / not same Impact Signals text · 738.0s
Official demo text from model authors; use for voice/style only, not final transcript-fidelity certainty.
https://microsoft.github.io/VibeVoice/assets/audio/3p_gpt5.mp3

VoxCPM2

Apache-2.0Commercial OK3 clips
Role: Designed voice without cloning / controllable speaker candidate
Evidence: Official public demo samples downloaded from OpenBMB demo page
Recommendation: Promising if local runtime can render stable designed English voices.
English voice design sample
download failed · unknowns
<HTTPError 404: 'Not Found'>
https://openbmb.github.io/voxcpm2-demopage/audio/voice_design/output.wav
English language sample
official demo sample / not same Impact Signals text · 8.52s
Official demo text from model authors; use for voice/style only, not final transcript-fidelity certainty.
https://openbmb.github.io/voxcpm2-demopage/audio/language/en_us.mp3
Cross-lingual English reference output
official demo sample / not same Impact Signals text · 13.6s
Official demo text from model authors; use for voice/style only, not final transcript-fidelity certainty.
https://openbmb.github.io/voxcpm2-demopage/audio/cross_lingual/en_en.wav

FireRedTTS2

Apache-2.0Commercial OK3 clips
Role: Dialogue-style / podcast generation candidate
Evidence: Official public demo samples downloaded from FireRedTTS2 demo page
Recommendation: Interesting for dialogue; must local-render Impact Signals text before promotion.
Podcast generation sample 1
download failed · unknowns
<HTTPError 404: 'Not Found'>
https://fireredteam.github.io/demos/firered_tts_2/audios/zero-shot-podcast-generation_samples/en_1_mooncast.wav
Podcast generation sample 2
download failed · unknowns
<HTTPError 404: 'Not Found'>
https://fireredteam.github.io/demos/firered_tts_2/audios/zero-shot-podcast-generation_samples/en_2_mooncast.wav
English zero-shot output sample
download failed · unknowns
<HTTPError 404: 'Not Found'>
https://fireredteam.github.io/demos/firered_tts_2/audios/ZeroShotICL_samples/mandarin_to_english/26900001%23en_024.wav

Fun-CosyVoice3

Apache-2.0Commercial OK3 clips
Role: Possible replacement for problematic CosyVoice2 path
Evidence: Official public demo samples downloaded from FunAudioLLM CosyVoice3 page
Recommendation: Revisit only if local CosyVoice3 passes WER; demo page provides C2 vs C3 comparison.
CosyVoice3 base English zero-shot
official demo sample / not same Impact Signals text · 4.92s
Official demo text from model authors; use for voice/style only, not final transcript-fidelity certainty.
https://funaudiollm.github.io/cosyvoice3/audio/c3_base/zero-shot/en/uttid_2.wav
CosyVoice3 large English zero-shot
official demo sample / not same Impact Signals text · 5.28s
Official demo text from model authors; use for voice/style only, not final transcript-fidelity certainty.
https://funaudiollm.github.io/cosyvoice3/audio/c3_large/zero-shot/en/uttid_2.wav
CosyVoice2 base English comparator
official demo sample / not same Impact Signals text · 5.96s
Official demo text from model authors; use for voice/style only, not final transcript-fidelity certainty.
https://funaudiollm.github.io/cosyvoice3/audio/c2_base/zero-shot/en/uttid_2.wav

Decision rules

  1. Use Kokoro samples for immediate production replacement decisions because they are local same-text renders.
  2. Use VibeVoice/VoxCPM2/FireRedTTS2/Fun-CosyVoice3 official demos to decide what deserves local install/render next.
  3. Do not promote a model until it renders Impact Signals sample text locally and passes WER/duration guardrails.
  4. All five candidates here have permissive model licenses for commercial podcast use, but cloned/reference voices still require independent rights.

Exact local sample text

Impact Signals is tracking how artificial intelligence is moving from demonstration projects into public health, disaster response, and humanitarian operations. The replacement voice should sound calm, credible, and clear. It should pronounce organizations, numbers, and policy terms without skipping words.