real-time speech translator on Deepgram and Groq

Open-Source Real-Time Speech Translator: A Build with Elixir, Rust, and Deepgram

Commercial real-time speech translation services fall short for seamless use in calls. Google Meet translates only 6 languages with a 2-3 second delay and requires a paid AI Pro subscription. JotMe, Talo, and Transync provide subtitles instead of voice, work only in Chrome or specific platforms like Zoom. Enterprise solutions like Palabra, KUDO, and Interprefy cost from $300/month and are geared toward conferences.

Physical devices miss the start of phrases, confuse words, and aren't suitable for live conversations. No solution offers voice translation in any application without ecosystem limitations.

| Solution | Price | Voice/Subtitles | Delay | Languages | Platforms |

|----------|-------|-----------------|-------|-----------|-----------|

| Google Meet | AI Pro Subscription | Voice | 2-3s | 6 | Meet only |

| JotMe | $10-15/month | Subtitles | ~1-2s | 77 | Chrome |

| Talo | Subscription | Subtitles + bot | ~2s | 60 | Zoom, Meet |

| Palabra | SaaS | Voice | ~0.8s | 25+ | Zoom, Meet |

| KUDO | $300+/month | Voice + humans | ~2-4s | 60+ | Proprietary |

Custom Pipeline: From STT to TTS

Capture audio from the microphone, recognize speech, translate, synthesize, and route to a virtual microphone. Reverse flow for the interlocutor's speech. Full v3 architecture with Elixir + Rust + Flask (single file), open-source.

Process:

• STT: Deepgram Nova-3 via WebSocket, 258-681ms, streaming with pauses.
• Translation: Groq with llama-3.3-70b, 250-560ms, prompt for literal translation with tone.
• TTS: Piper locally, 300-500ms, 29 languages offline.
• Audio: Virtual microphone/speaker for any application (Meet, Zoom, Discord).

Total delay 0.8-1.7s — faster than a live interpreter (2-5s).

Component Selection Based on Benchmarks

STT: Deepgram vs Alternatives

• voxtral.c: 15+s locally, glitches on Apple Silicon.
• Groq Whisper: ~500ms, garbage in chunks.
• Deepgram: Streaming, finalized text on pauses.

LLM Translation: Groq Leads

| Provider | Delay | Price | Drawbacks |

|----------|-------|-------|-----------|

| Google Translate | 100-300ms | $20/1M | Worse than LLM on conversational speech |

| DeepL | 200-500ms | $25/month | Subscription, worse than LLM |

| OpenAI | 500-1200ms | Paid | Slow |

| Groq | 250-560ms | Free | Optimal |

llama-3.3-70b on Groq's LPU chips — free tier for basic tasks.

TTS: Piper for Offline Use

• ElevenLabs: Excellent, but $5.57/hour.
• Cartesia: $1.26/hour.
• Piper: Acceptable quality, free, 29 languages.

Cost and Implementation

• Deepgram: $200 credits (hundreds of hours).
• Groq: Free.
• Piper: Offline.

Total ~$0. Testing on Meet/Zoom with a team showed: translation is accurate in meaning but loses idioms/sarcasm. Piper's voice sounds robotic (Russian 'Denis' — like a 1980s professor). Delay + network ping = noticeable but functional.

Installation: Elixir for orchestration, Rust for audio, Flask as a bridge. Works on macOS Apple Silicon.

Key Takeaways

• Full compatibility with any VoIP: capture at audio driver level.
• Delay of 0.8-1.7s thanks to Deepgram streaming + fast Groq.
• Free: $200 Deepgram credit for years, rest open-source/free.
• Trade-offs: Piper sounds robotic, LLM misses nuances of conversational speech.
• Open-source: repository for customization (Elixir/Rust stack).

— Editorial Team