Tts

Workflow

1. Engine Selection

• MOSS-TTS (GGUF): Best for stable, high-quality generation. Ideal for long-form content.
• OmniVoice: Best for rapid voice design using specific tags or zero-shot cloning.
• VoxCPM2: Best for GPU-first expressive synthesis and basic cloning when you want to test whether it can outperform MOSS on the local RTX 4050 6 GB setup.

2. Quick Selection Guide

• Use MOSS-TTS when stability matters most or the text is longer.
• Use OmniVoice when the user wants a voice described with explicit tags such as accent, gender, age, or pitch.
• Use VoxCPM2 when the user wants cloning from a real speech sample.
• When the user provides both a speech sample and its exact transcription, use the clone CLI path as the default, not as an optional upgrade.
• Do not ask for generic voice labels such as male calm or female assertive if a cloned voice is the goal.

3. Speech Generation

Using MOSS-TTS

python generate_moss.py --text "Your text" --output "out.wav" [--reference "ref.wav"]

Using OmniVoice

.\venv\Scripts\python.exe generate_omnivoice.py --text "Your text" --instruct "tags" --output "out.wav"

Using VoxCPM2

Plain TTS:

.\venv\Scripts\python.exe generate_voxcpm.py --text "Your text" --output "out.wav"

Basic cloning:

.\venv\Scripts\python.exe generate_voxcpm.py --text "Your text" --reference "ref.wav" --output "out.wav"

Default cloning path when the user has a speech sample and exact transcription:

.\venv\Scripts\python.exe -m voxcpm.cli clone ^
  --text "Final target text" ^
  --prompt-audio "voice_sample.wav" ^
  --prompt-text "Exact transcript of voice_sample.wav" ^
  --reference-audio "voice_sample.wav" ^
  --output "out.wav"

Use this mode by default for cloning whenever the user can provide the exact transcript of the sample audio. This was the best-performing local workflow for Polish cloning in practice.

Preferred PowerShell invocation for reliable local runs:

$text = @'
Final target text
'@
$prompt = @'
Exact transcript of the sample audio
'@
& 'C:\path\to\venv\Scripts\python.exe' -m voxcpm.cli clone `
  --text $text `
  --prompt-audio 'C:\path\to\voice_sample.wav' `
  --prompt-text $prompt `
  --reference-audio 'C:\path\to\voice_sample.wav' `
  --output 'C:\path\to\out.wav'

Use this PowerShell here-string pattern when the text is multiline, contains punctuation-heavy Polish narration, or when reproducing a known-good local command exactly matters more than using intermediate text files.

Optional micro-cue pattern for cloned speech:

$text = @'
(gentle wrap-up)Final target text
'@

Use this only as a very soft, short parenthetical cue at the start of the text when the caller explicitly wants boundary-sensitive prosody nudges. Treat it as a subtle delivery hint, not full voice design.

Prepared sample commands without execution:

.\venv\Scripts\python.exe generate_voxcpm_samples.py

Valid English Tags (Strict): american accent, australian accent, british accent, canadian accent, chinese accent, indian accent, japanese accent, korean accent, portuguese accent, russian accent, female, male, child, teenager, young adult, middle-aged, elderly, very high pitch, high pitch, moderate pitch, low pitch, very low pitch, whisper.

4. VoxCPM2 Practical Guidance

• The local wrapper generate_voxcpm.py only exposes plain TTS and basic --reference cloning.
• For cloning, prefer the upstream CLI entrypoint: .\venv\Scripts\python.exe -m voxcpm.cli ...
• Treat prompt-audio + prompt-text + reference-audio as the default clone path when the user can provide a transcript of the sample.
• prompt-text should be the exact spoken words from the sample audio, not a paraphrase.
• If the user wants cloning and does not provide a transcript, ask for it explicitly before continuing.
• Only fall back to basic --reference cloning if the user explicitly approves that downgrade.
• Avoid adding style instructions if the goal is to preserve the sample's original speaking style as closely as possible.
• When reproducing a known-good clone command, do not add extra tuning flags unless the user explicitly asks for them.
• For the default voxcpm.cli clone path, do not add --normalize, --control, --no-optimize, custom --cfg-value, or custom --inference-timesteps unless the user explicitly requests experimentation.
• If the caller supplies a parenthetical cue at the start of the text, treat it as a micro-prosody nudge only.
• Keep such cues very short and sparse so the model stays close to the cloned speaker identity.
• Prefer cues about discourse position or transition, such as clear start, steady continuation, slight contrast, or gentle wrap-up, over strong mood or persona descriptions.
• If a cue makes the output sound less like the speaker, remove it rather than strengthening it.
• Prefer handling text normalization at the TTS stage rather than pushing engine-specific rewrites upstream into the scriptwriter skill.
• If the narration contains digits, dates, abbreviations, passwords, or mixed-language tokens and the raw clone sounds garbled, test --normalize first so wetext can expand the text automatically.
• Only manually rewrite the target text when wetext normalization is unavailable, insufficient, or the user explicitly wants hand-normalized phrasing.
• Use the local low-VRAM defaults first when staying in the wrapper: optimize=False, load_denoiser=False, cfg_value=2.0, inference_timesteps=8.
• Plain English and Polish TTS were verified to fit in VRAM at roughly 5247 MiB allocated and ~5.6-5.7 GiB reserved.
• Start with plain TTS before cloning. Cloning is the next likely point of VRAM failure.

5. Notes

• VoxCPM2 uses the installed voxcpm package and the local wrapper generate_voxcpm.py.
• The wrapper is convenient, but it does not expose the stronger clone workflow that uses prompt audio and a transcript.
• When the user wants a cloned voice, use python -m voxcpm.cli clone instead of the wrapper unless they explicitly ask for a fallback.
• VoxCPM2 basic clone mode uses --reference and maps to the upstream reference_wav_path API.
• generate_voxcpm_samples.py prepares the numbered sample set voxcpm_1.wav, voxcpm_2.wav, voxcpm_3.wav, and voxcpm_clone_test.wav.
• The clone sample expects a placeholder reference file at temp\voxcpm_clone_reference.wav.

References

• Config: MOSS-TTS uses moss_local_config.yaml.
• Models:
  • MOSS: models/MOSS-TTS-GGUF/MOSS_TTS_Q4_K_M.gguf
  • OmniVoice: drbaph/OmniVoice-bf16
  • VoxCPM2: openbmb/VoxCPM2