feat: add ElevenLabs Scribe STT transcriber and Telegram SendVoice support

[manaporkun]

Summary

• Add ElevenLabsTranscriber as an alternative speech-to-text provider using the ElevenLabs Scribe API (scribe_v1). This enables voice message transcription for users who already have an ElevenLabs API key, without requiring a separate Groq account.
• Add Telegram SendVoice support so voice messages are sent as proper voice bubbles instead of audio file attachments.

Changes

File	Change
pkg/voice/elevenlabs_transcriber.go	New ElevenLabsTranscriber struct implementing the Transcriber interface. Uses xi-api-key header auth and scribe_v1 model.
pkg/voice/transcriber.go	Updated DetectTranscriber to check voice.elevenlabs_api_key first, falling back to Groq model-list entries for backward compatibility.
pkg/config/config.go	Added ElevenLabsAPIKey field to VoiceConfig.
pkg/channels/telegram/telegram.go	Telegram-specific voice-bubble detection inside "audio" case — OGG files with "voice" in filename use SendVoice, others use SendAudio. No changes to inferMediaType or other channels.
pkg/voice/elevenlabs_transcriber_test.go	TestElevenLabsTranscribe suite (success, API error, missing file).
pkg/voice/transcriber_test.go	DetectTranscriber priority tests for ElevenLabs, Groq, and voice model name.

Configuration

{
  "voice": {
    "elevenlabs_api_key": "sk_your_key_here"
  }
}

Or via environment variable: PICOCLAW_VOICE_ELEVENLABS_API_KEY=sk_your_key_here

When configured, ElevenLabs takes priority over Groq for transcription. Existing Groq configurations (via model_list) continue to work unchanged.

Test plan

• All existing tests pass
• New ElevenLabs transcriber tests pass (success, API error, missing file)
• DetectTranscriber correctly prioritizes: voice model name > ElevenLabs > Groq model-list
• Interface compliance verified at compile time
• Tested on Raspberry Pi Zero 2 W with real ElevenLabs API
• Telegram voice bubbles render correctly on iOS/Android

[CLAassistant]

All committers have signed the CLA.

[Copilot]

Pull request overview

This PR adds an ElevenLabs Scribe-based speech-to-text transcriber option and improves Telegram outbound media handling by sending OGG “voice note” media as Telegram voice bubbles.

Changes:

• Added ElevenLabsTranscriber (Scribe STT) and updated DetectTranscriber to prefer ElevenLabs when configured.
• Added Telegram outbound "voice" handling via SendVoice.
• Introduced filename-based "voice" media type inference for OGG/OGA files containing "voice" in the filename.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
pkg/voice/transcriber.go	Adds ElevenLabs STT transcriber and updates provider detection priority.
pkg/voice/transcriber_test.go	Adds interface compliance + ElevenLabs transcribe tests and provider priority tests.
pkg/config/config.go	Extends ProvidersConfig with ElevenLabs and updates IsEmpty().
pkg/agent/loop.go	Adds "voice" inference in inferMediaType based on filename for OGG/OGA.
pkg/channels/telegram/telegram.go	Adds "voice" branch to send Telegram voice bubbles via SendVoice.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

inferMediaType now returns "voice" for files named like "voice.ogg". This value is propagated into bus.MediaPart.Type for all channels (see where tool media parts are built), but only Telegram has been updated to handle "voice". Channels that map types explicitly (e.g., OneBot maps only "audio" -> "record", WeCom maps only "audio" -> "voice") will now treat these as generic files, which is a functional regression for non-Telegram channels. Consider keeping the inferred type as "audio" and doing Telegram-specific voice-bubble detection inside the Telegram channel, or ensure every channel that handles audio also treats "voice" as audio-equivalent.

Suggested change

	// Detect voice messages: OGG files with "voice" in the filename.
	// These are sent as Telegram voice bubbles rather than audio attachments.
	if strings.Contains(fn, "voice") && (strings.HasSuffix(fn, ".ogg") || strings.HasSuffix(fn, ".oga")) {
	return "voice"
	}

[manaporkun]

Good catch — fixed. Moved voice-bubble detection out of inferMediaType (which now always returns "audio" for OGG files) and into the Telegram channel's Send method. Other channels are unaffected.

[afjcjsbx]

a new branch was mergiated, please could you fix the conflicts? 🙏

[manaporkun]

Rebased on latest main and resolved conflicts. Adapted to the new transcriber architecture — ElevenLabs transcriber is now in its own file (elevenlabs_transcriber.go) matching the new pattern, and DetectTranscriber checks voice model name first, then ElevenLabs, then Groq. All 25 tests pass.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The voice-bubble detection uses strings.Contains(fn, "voice"), which will also match unrelated filenames like invoice.ogg and send them as Telegram voice messages incorrectly. Consider using a stricter check (e.g., HasPrefix/HasSuffix on the base name with a delimiter, or a regex for (^|[^a-z0-9])voice([^a-z0-9]|$)) so only intended voice clips are routed to SendVoice.

[Copilot]

On non-200 responses, the code logs and returns the full response body ("response": string(body) and includes it in the returned error). This can leak potentially sensitive information and can also produce very large logs/errors if the upstream returns HTML or verbose JSON. Consider truncating/sanitizing the body in logs/errors (and optionally parsing a structured error field) while still preserving enough detail for debugging.

Suggested change

	logger.ErrorCF("voice", "ElevenLabs API error", map[string]any{
	"status_code": resp.StatusCode,
	"response": string(body),
	})
	return nil, fmt.Errorf("ElevenLabs API error (status %d): %s", resp.StatusCode, string(body))
	truncatedBody := utils.Truncate(string(body), 512)
	logger.ErrorCF("voice", "ElevenLabs API error", map[string]any{
	"status_code": resp.StatusCode,
	"response": truncatedBody,
	})
	return nil, fmt.Errorf("ElevenLabs API error (status %d): %s", resp.StatusCode, truncatedBody)

[afjcjsbx]

LGTM!

[afjcjsbx]

pls fix lint 🙏

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The PR description/config snippet refer to providers.elevenlabs.api_key, but the implementation adds voice.elevenlabs_api_key (and DetectTranscriber reads cfg.Voice.ElevenLabsAPIKey). This mismatch will cause users following the documented JSON to get no ElevenLabs transcriber. Either wire ElevenLabs through providers.elevenlabs as described, or update the docs/PR description and any config examples to match the voice section.

[manaporkun]

Good catch! The original implementation used providers.elevenlabs, but during rebase onto latest main, the ProvidersConfig struct was removed (provider refactoring to model_list). I moved the ElevenLabs API key to voice.elevenlabs_api_key instead, which fits better since it's specifically for voice transcription. Updated the PR description and config examples to match the actual implementation.

[huaaudio]

Hi @manaporkun , thanks for the PR! This config logic is pretty good. Can you run make fmt && make lint locally to fix the lint issue before we proceed and merge? Thanks

[Copilot]

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated no new comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Orgmar]

@manaporkun Nice contribution! The ElevenLabs Scribe transcriber gives users another solid STT option, and the priority chain (voice model > ElevenLabs > Groq) keeps backward compatibility clean. The Telegram SendVoice bubble support is a great UX improvement too, way better than raw audio attachments. Thorough test coverage across the board.

We're running a PicoClaw Dev Group on Discord for contributors to chat and collaborate. If you're interested, email support@sipeed.com with subject [Join PicoClaw Dev Group] manaporkun and we'll get you the invite!