LibreChat voice: hosted Whisper and text to speech in librechat.yaml
LibreChat's OpenAI speech blocks take a custom URL. Two blocks in librechat.yaml give your users voice input and read-aloud without running a speech server.
1. Get a key
Get an access key on the pricing page: $1 once (1,000 credits) or $3 a month (3,000 credits, about 60 hours). After paying you get a page that shows the key (it starts with adt_). That key is the API key in every app below.
2. Put the key in .env
STT_API_KEY=adt_your_key
TTS_API_KEY=adt_your_key
3. Add the speech block to librechat.yaml
speech:
stt:
openai:
url: 'https://glyphcall.com/v1/audio/transcriptions'
apiKey: '${STT_API_KEY}'
model: 'whisper-1'
tts:
openai:
url: 'https://glyphcall.com/v1/audio/speech'
apiKey: '${TTS_API_KEY}'
model: 'tts-1'
voices: ['nova', 'alloy', 'onyx', 'shimmer', 'echo', 'fable']
maxInputChars: 2000
LibreChat wants the full endpoint path in url, not the base. maxInputChars: 2000 matters: LibreChat sends up to 20,000 characters by default, and one speech request here takes at most 2,000 (a longer one is refused, uncharged). If your LibreChat version does not accept maxInputChars in that place, check its speech configuration page for where it goes.
Restart LibreChat, then pick the OpenAI engine in each user's Speech settings.
Test the key first
If something does not work in the app, check the key and the endpoint with one call from a terminal:
curl https://glyphcall.com/v1/audio/speech \
-H "Authorization: Bearer YOUR_KEY" -H "Content-Type: application/json" \
-d '{"model":"tts-1","input":"Hello from Glyphcall Voice.","voice":"nova"}' -o hello.mp3
What it costs
| Call | Price |
|---|---|
| Text to speech, up to 300 characters | $0.01 (10 credits) |
| Text to speech, 301 to 2,000 characters | $0.08 (80 credits) |
| Speech to text, a file up to 5 MB | $0.01 (10 credits) |
| Speech to text, a larger MP3 or WAV up to 60 minutes / 25 MB | $0.05 (50 credits) |
A failed call is never charged. No account, no subscription unless you choose the monthly plan.
Questions
Which voice names can I list? OpenAI's names and the 40 English voice names on the voice list. They all work in the voices array.
Does it need allowedAddresses? No. That list is for private hosts; this endpoint is public.
Other guides: Open WebUI · AnythingLLM · Home Assistant · Obsidian Whisper · Any other app · OpenAI tts-1 shutdown · OpenAI whisper-1 shutdown · Voice list