You are viewing the documentation for a prerelease version. View Latest

Synthesizing speech

Dengjen synthesizes speech from a Piper voice. Download a voice’s .onnx model and matching .onnx.json config from the Piper voices repository and keep both files together, e.g. voices/en_US-lessac-medium.onnx and voices/en_US-lessac-medium.onnx.json.

A voice doesn’t have to come from the official Piper voices repository — any VITS-family ONNX export using the same 3/4-input tensor convention (phoneme ids, lengths, scales, optional speaker id) can be loaded by writing a matching .onnx.json manifest with "model_type": "vits". The minimal required fields are audio (sample rate), inference (noise_scale/length_scale/noise_w), and phoneme_id_map (the symbol vocabulary); phoneme_type selects the phonemizer (espeak is the default if omitted, and needs an espeak.voice entry — text, hebrew, and pinyin don’t).

A MeloTTS voice — its own VITS-derived ONNX export, with a tones tensor alongside phone ids — is loaded with "model_type": "melotts". The required fields are audio (sample rate), inference (noise_scale/length_scale/noise_scale_w), phone_id_map and tone_id_map (the phone and tone symbol vocabularies), and model_path; phonemizer selects the phonemization backend and must be one of {"type": "espeak", "voice": "<espeak-ng voice name>"} (covers English, Spanish, French, Japanese, Korean) or {"type": "pinyin", "model_dir": "<g2pW model directory>"} (Chinese, with real tone extraction — requires building with the pinyin feature).

CLI

Synthesize text from a file to a WAV file, using the dengjen-cli frontend:

cargo run --release -p dengjen-tts-cli -- voices/en_US-lessac-medium.onnx.json \
    -f input.txt \
    -o output.wav

Or send a single request as JSON on stdin and capture the WAV bytes written to stdout:

echo '{"text": "Hello world"}' | cargo run --release -p dengjen-tts-cli -- voices/en_US-lessac-medium.onnx.json > output.wav

Run dengjen --help (or cargo run --release -p dengjen-tts-cli — --help) for the full list of options. The main ones:

--mode lazy|parallel|batched|realtime selects the synthesis mode (default lazy). --batch-size sets how many sentences batched synthesizes per concurrent group, and --chunk-size and --chunk-padding tune realtime chunking.
--speaker-id picks a speaker for multi-speaker voices; when omitted, the voice’s own default speaker from its manifest is used (Kokoro’s is 0).
--rate, --pitch and --volume take 0-100. When omitted, each stays at 1.0 (no change). The value maps linearly onto a speed of 0.5x-5.5x (so 10 is 1x and 50 is 3x), a pitch of 0.5-1.5 and a volume gain of 0-1 (a volume of 0 mutes the audio). --silence appends that many milliseconds of silence after each sentence; --rate also changes how long that silence lasts, because the silence is processed like the rest of the audio.
--length-scale, --noise-scale and --noise-w override inference knobs: --noise-w is Piper’s, while --length-scale and --noise-scale also apply to MeloTTS. --param KEY=VALUE (repeatable) sets a named backend parameter to a number, such as MeloTTS’s noise_scale_w (an unrecognized key may be silently ignored); a named flag wins over a conflicting --param.

Want to help? Learn how to contribute to the ZirekHQ docs ›