Observing the VOICEVOX API
VOICEVOX is stated to be composed of an editor, an engine, and a core.
Reference: Overall Structure
It seems the editor is the application, the engine is an HTTP server, and the core is a module that performs speech synthesis processing.
This implies that the editor makes REST API calls (hereinafter referred to as API) to the engine.
So, this article will observe the content of that API.
Wireshark was used to capture the API traffic.
Communication on Startup
Here are the results after filtering with http and tcp.port == 50021.
The following information appears to be read on startup:
- Version information
/version - Engine manifest information
/engine_manifest - Speaker information
/speakers(character list like Zundamon) - Singer information
/singers(same as above)
After obtaining speaker/singer information, more detailed information for each character is retrieved (e.g., /speaker_info?speaker_uuid=xxx, /singer_info?speaker_uuid=xxx).
Communication during Speech Synthesis Request
Now, I sent a speech synthesis request with Zundamon and peeked at the API.
It seems that audio is acquired in the following flow:
- Accent information via
/accent_phrases - Speech synthesis of Zundamon's voice via
/synthesis?speaker=3
The request body sent in (2.) is similar to the response from (1.).
Therefore, the flow appears to be: get accents in (1.), then synthesize speech from those accents in (2.).
Actually Calling the API
I used the httpie tool to call the API.
- Get Speaker Information
It was found that Zundamon (Normal) has an ID of 3.
- Get Accent Information
I tried to get accent information for ずんだもんなのだ (Zundamon nanoda).
(Unlike speaker information, this is retrieved with a POST request.)
- Speech Synthesis
Create a request body like the following:
{
"accent_phrases": </data obtained from /accent_phrases>,
"speedScale": 1,
"pitchScale": 0,
"intonationScale": 1,
"volumeScale": 1,
"prePhonemeLength": 0.1,
"postPhonemeLength": 0.1,
"outputSamplingRate": 24000,
"outputStereo": false,
"kana": ""
}
Since httpie cannot handle WAV files, I will send the request using PowerShell.
# Define URL and JSON data
$url = 'http://localhost:50021/synthesis?speaker=3'
$jsonBody = @"
{
"accent_phrases": [
{
"moras": [
{
"text": "ズ",
"consonant": "z",
"consonant_length": 0.12722788751125336,
"vowel": "u",
"vowel_length": 0.11318323761224747,
"pitch": 5.773037910461426
},
{
"text": "ン",
"consonant": null,
"consonant_length": null,
"vowel": "N",
"vowel_length": 0.09306197613477707,
"pitch": 6.108947277069092
},
{
"text": "ダ",
"consonant": "d",
"consonant_length": 0.04249810427427292,
"vowel": "a",
"vowel_length": 0.09372275322675705,
"pitch": 6.09743070602417
},
{
"text": "モ",
"consonant": "m",
"consonant_length": 0.07012023776769638,
"vowel": "o",
"vowel_length": 0.1172478124499321,
"pitch": 5.932623386383057
},
{
"text": "ン",
"consonant": null,
"consonant_length": null,
"vowel": "N",
"vowel_length": 0.06496299058198929,
"pitch": 5.745952129364014
},
{
"text": "ナ",
"consonant": "n",
"consonant_length": 0.038462959229946136,
"vowel": "a",
"vowel_length": 0.08576127141714096,
"pitch": 5.5794854164123535
}
],
"accent": 1,
"pause_mora": null,
"is_interrogative": false
},
{
"moras": [
{
"text": "ノ",
"consonant": "n",
"consonant_length": 0.05504273623228073,
"vowel": "o",
"vowel_length": 0.0903041884303093,
"pitch": 5.551316261291504
},
{
"text": "ダ",
"consonant": "d",
"consonant_length": 0.05024997144937515,
"vowel": "a",
"vowel_length": 0.20450790226459503,
"pitch": 5.633930206298828
}
],
"accent": 2,
"pause_mora": null,
"is_interrogative": false
}
],
"speedScale": 1,
"pitchScale": 0,
"intonationScale": 1,
"volumeScale": 1,
"prePhonemeLength": 0.1,
"postPhonemeLength": 0.1,
"outputSamplingRate": 24000,
"outputStereo": false,
"kana": ""
}
"@
# Create HTTP headers
$headers = @{
'Content-Type' = 'application/json'
}
# Send POST request and get response
$response = Invoke-WebRequest -Uri $url -Method Post -Headers $headers -Body $jsonBody -OutFile "output.wav"
# Open and play
start output.wav
VOICEVOX: Zundamon
That's all!
