Voice Transcription

1. What? — Definition and context

The Voice Transcription workflow automatically transcribes Telegram voice messages into text. A smart routing system picks the optimal transcription service based on the message duration.

Whisper

Whisper is an AI model developed by OpenAI that converts speech to text. Groq offers an ultra-fast and free version, perfect for short messages. ElevenLabs Scribe handles long files with diarisation (speaker identification).

Services used

Service	Usage	Advantage
Groq Whisper	Messages ≤ 30s	Free, fast (1-2s)
ElevenLabs Scribe	Messages > 30s	Diarisation, long files

Routing by duration

Duration	Service	Reason
≤ 30 seconds	Groq Whisper	Fast, free
> 30 seconds	ElevenLabs Scribe	Diarisation, long files

---

2. Why? — Stakes and motivations

Problems solved

Problem	Without transcription	With transcription
Mandatory listening	Replay to understand	Readable text instantly
No search	No ctrl+F on audio	Indexable text
Hard to share	Send the audio file	Copy-paste the text
Accessibility	Not accessible to deaf users	Universal text

Why two services?

Criterion	Groq Whisper	ElevenLabs Scribe
Cost	Free	Paid (per hour)
Speed	~1-2s	~10-30s
File limit	25 MB	3 GB
Diarisation	No	Yes
Best for	Short messages	Meetings, podcasts

Why 30 seconds?

Groq Whisper is optimised for short files and returns the transcription in under 2 seconds. For longer files, ElevenLabs offers better quality with speaker identification.

---

3. How? — Technical implementation

Architecture

Workflow input

{
  "message": {
    "voice": {
      "file_id": "AwACAgIAAxkB...",
      "duration": 15,
      "mime_type": "audio/ogg"
    },
    "from": {
      "id": 123456789,
      "first_name": "Guillaume"
    },
    "chat": {
      "id": 123456789
    }
  }
}

Groq Whisper configuration (≤30s)

Community Node: n8n-nodes-groq

Parameter	Value
Credential	Groq account - N8N
Operation	Transcribe
Model	whisper-large-v3-turbo
Input Data Field	data
Language	fr (optional)
Response Format	json

ElevenLabs Scribe configuration (>30s)

HTTP Request Node

Parameter	Value
Method	POST
URL	https://api.elevenlabs.io/v1/speech-to-text
Authentication	Header Auth → ElevenLabs API
Body Content Type	Form-Data

Form Parameters:

Name	Type	Value
file	Binary	{{ $binary.data }}
model_id	String	scribe_v1
language_code	String	fr

Output

{
  "success": true,
  "text": "Remind me to call Jean tomorrow",
  "duration": 15,
  "service": "groq"
}

Integration with the orchestrator

The Telegram Orchestrator detects voice notes and calls this sub-workflow:

IF message.voice exists:
  Execute Workflow: Voice Transcription
  Input: $json (contains message.voice)

  IF response.success:
    IF active conversation exists (#231/#232):
      Send Transcript Preview + route text to Conversation Agent
    ELSE:
      Send message: "🎤 {response.text}"
  ELSE:
    Send message: "❌ Transcription failed"

Since Phase 5 (#231/#232), if a conversation is active when a voice note arrives, the transcribed text is injected as a message into the conversation instead of being returned as-is to the user. This enables a voice discussion with the bot.

Post-transcription callbacks

Callback	Action
voice_retry_{msg_id}	Retry transcription
voice_process_{msg_id}	Process with Claude (summary, extraction)
voice_save_{msg_id}	Save as a note

---

4. What if? — Outlook and limits

Limits and costs

Service	File limit	Cost	Speed
Groq Whisper	25 MB	Free	~1-2s
ElevenLabs Scribe	3 GB	Paid (per hour)	~10-30s

Groq quotas

Groq has per-minute request limits. For usage spikes, the workflow automatically falls back to ElevenLabs.

Current limits

Limit	Impact	Mitigation
Groq quota	Possible rate limiting	ElevenLabs fallback
OGG format	Telegram-only format	APIs natively supported
No diarisation < 30s	No speaker identification	Acceptable for short messages

Evolution scenarios

If Groq rate limit is hit:

• Temporarily set duration ≤ 0 to force ElevenLabs
• Or add OpenAI Whisper as an intermediate fallback

If systematic diarisation is needed:

• Route every message to ElevenLabs
• Or use a local model with speaker detection

If multi-language is needed:

• Auto-detect the language
• Adapt parameters according to detected language

Troubleshooting

Problem	Check
Empty transcription	Does the audio file actually contain speech?
ElevenLabs timeout	Files > 5min: increase timeout (180s)
Groq rate limit	Check quotas, fall back to ElevenLabs
Unsupported format	Telegram sends .ogg (Opus) — natively supported

---

Related pages

Workflows

• Telegram Orchestrator — Central hub
• Notification Hub — Notification routing

Infrastructure

• AI Stack — Claude Ollama for post-processing

External references

• Groq Speech-to-Text
• ElevenLabs STT API