Providers

Talkative Lobster supports multiple speech-to-text (STT) and text-to-speech (TTS) providers. Choose the combination that best fits your needs.

Speech-to-Text (STT)

Provider	Type	Languages	API Key	Latency
ElevenLabs Scribe	Cloud	Multilingual	Required	Low
OpenAI Whisper	Cloud	Multilingual	Required	Low
whisper.cpp	Local	Multilingual	Not needed	Medium

ElevenLabs Scribe

High-accuracy cloud STT using ElevenLabs' Scribe v2 model.

• API key: Get one at elevenlabs.io
• Model: scribe_v2 (fixed)
• Timeout: 5 seconds
• Best for: Production-quality multilingual transcription

OpenAI Whisper

OpenAI's cloud-hosted Whisper model (whisper-1).

• API key: Get one at platform.openai.com
• Timeout: 5 seconds
• Best for: General-purpose transcription

whisper.cpp

Runs Whisper locally on your machine. No data leaves your device.

• Binary: You need the whisper-cli binary — build from whisper.cpp or install via Homebrew: brew install whisper-cpp
• Model: ggml-medium.bin — automatically downloaded to ~/.config/lobster/models/ on first use
• Language: Hardcoded to Japanese (--language ja)
• Timeout: 60 seconds
• Best for: Privacy-conscious use, offline operation

STT Fallback

If multiple providers are configured with valid keys, the app tries providers in order (ElevenLabs → OpenAI → whisper.cpp) and uses the first successful result.

---

Text-to-Speech (TTS)

Provider	Type	Languages	API Key	Streaming
ElevenLabs	Cloud	Multilingual	Required	Yes
VOICEVOX	Local	Japanese	Not needed	No
Kokoro	Local	Japanese, English	Not needed	No
Piper	Local	Many	Not needed	No

ElevenLabs

High-quality cloud TTS with natural-sounding voices and real-time streaming.

• API key: Same as ElevenLabs Scribe (ELEVENLABS_API_KEY)
• Output format: PCM 24kHz, 16-bit mono (streamed)

Voices

Voice	ID
Morioki	KnMBELSmBGHPqfZxMRw6
Lily (default)	pFZP5JQG7iQjIQuC4Bku
Alice	Xb7hH8MSUJpSbSDYk0k2
Matilda	XrExE9yKIg1WjnnlVkGX
Sarah	EXAVITQu4vr4xnSDxMaL
Daniel	onwK4e9ZLuTAKqWW03F9
Brian	nPczCjzI2devNBz1zQrb
George	JBFqnCBsd6RMkjVDRZzb
Liam	TX3LPaxmHKxFdv7VOQHJ

Models

Model	Description
eleven_multilingual_v2	Highest quality, multilingual (default)
eleven_turbo_v2_5	Balanced quality and speed
eleven_flash_v2_5	Fastest response time

VOICEVOX

Free, open-source Japanese TTS engine. Runs as a local HTTP server.

• Download: voicevox.hiroshiba.jp
• Server URL: Default http://localhost:50021
• Speaker ID: Integer (default: 1). See VOICEVOX docs for available speakers.
• Process: 2-step — audio_query then synthesis
• Output format: WAV

WARNING

VOICEVOX must be running before you start TalkLob. The app connects to its HTTP API.

Kokoro

Lightweight local TTS supporting Japanese and English.

• Server URL: Default http://localhost:8880
• API: OpenAI-compatible (POST /v1/audio/speech)
• Output format: MP3

Voices

Voice	Language
jf_alpha (default)	Japanese
jf_gongitsune	Japanese
jf_nezumi	Japanese
jf_tebukuro	Japanese
jm_kumo	Japanese
af_heart	English
af_jadzia	English
af_jessica	English

Piper

Fast local TTS with broad language support. Runs entirely on your machine as a subprocess.

• Binary: Download from Piper releases
• Model: ONNX voice model file (.onnx)
• Timeout: 30 seconds per synthesis
• Output format: WAV

You need to set both the binary path and model path in Settings.