Skip to main content

​
Overview

GiggySTTService streams live input audio through Giggy’s paid transcription API. It emits replaceable interim transcript snapshots during capture and one final transcript after verifying durable completion. Developers keep their existing language model, context aggregators and transport, and may use Giggy TTS for output. The integration is maintained by Giggy. It requires upstream VAD, mono PCM16 LE at 16 kHz and automatic language detection. It does not supply word timestamps, speaker diarization, audio resampling or server-side turn detection.

​
Installation and compatibility

Install the Giggy-maintained integration from PyPI:
Requires Python 3.11 or later. Version 0.2.0 is validated with Python 3.12 and Pipecat 1.12.0; the dependency remains pinned to that Pipecat version. The live wire format is verified against Giggy’s current V17 admission and microphone-input handlers using offline worker/storage stand-ins. Admission uses JSON mode:"streaming", microphone:true. The adapter supplies a fresh valid request UUID per utterance; V17 also accepts omission and generates the UUID. Neither form resumes a disconnected session. Readiness supplies the admitted UUID, attempt and input capability used for ordered PCM frame POSTs.

​
Prerequisites

  • Create a Giggy account and your own API key.
  • Add paid Streaming credits. See Giggy pricing.
  • Keep GIGGY_API_KEY on your application server. STT needs no voice UUID.
  • Configure transport input as mono PCM16 at 16 kHz.
  • Supply VADUserStartedSpeakingFrame and VADUserStoppedSpeakingFrame upstream of STT, with total VAD onset/detection delay no greater than one second.
See the Giggy Speech API reference for the live PCM protocol and account/key setup.

​
Usage

The application owns and closes the session. Do not configure default authentication headers on it. Giggy TTS output is 24 kHz, independently of the 16 kHz STT input.

​
Transcript handling and utterance limits

Each InterimTranscriptionFrame contains a complete provisional snapshot. Replace displayed text; do not append revisions or commit them to LLM context. Only the canonical artifact becomes a finalized TranscriptionFrame. Each VAD-delimited captured utterance, including real-audio pre-roll and stop silence, must contain 0.5–30 seconds of audio. Shorter segments report an error before paid admission. Longer segments fail and close the stream. No silence is fabricated and no automatic rollover, reconnect or retry occurs. One previous utterance may finalize while one new utterance is captured. Network operations and final transcripts stay ordered. Slow finalization delays the next admission; a third pending utterance fails visibly rather than accumulating an unbounded queue. Each utterance is a separate paid Streaming operation, using Giggy’s existing actual-duration settlement and cancellation/refund contract.

​
Cancellation and errors

InterruptionFrame passes through without cancelling transcription of the user’s new speech. The application’s output transport handles bot playback interruption. CancelFrame, mute and cleanup close unfinished STT streams and sender tasks. EndFrame flushes capture and waits for completion. Readiness identity/attempt, transcript ordering, input acknowledgements, completed status and artifact duration are validated. Failed or ambiguous requests are never retried. The adapter does not switch to completed-WAV uploads or another STT provider. Errors omit API keys, input capabilities and provider response bodies. No measured STT latency guarantee is claimed.

​
Example and demo

The source repository contains examples/transcribe.py, replaying one existing 0.5–30 second mono PCM16 16 kHz WAV at capture cadence through the live input protocol. It requires a Giggy key, paid Streaming credits and GIGGY_ALLOW_BILLABLE_EXAMPLE=1; no LLM or TTS is needed. Watch the two-voice STT conversation demo. It shows real interim/final transcription, repeated turns and VAD-driven interruption of assistant playback. User utterances are prerecorded Giggy TTS; assistant replies are scripted and synthesized with Giggy TTS. Displayed assistant text timing is approximate. This recording does not qualify physical microphone hardware. Production recorded-input conversation and separate cancellation/refund checks passed.