My Simple Speech To Text

The Zero-UI Transcription Stack: Native-feel on iOS/Mac, serverless-fast, nearly free

This project turns S3 into your “native” transcription workspace. Record on iPhone, tap Share → “Upload to S3,” and your text shows up moments later in the same Files view—on iOS and macOS—without opening an app. Under the hood: an S3 event triggers a Chalice-powered Lambda, Whisper transcribes, Gemini tidies to clean Markdown, the result is saved right back to S3, and your phone gets a push notification.

Why this is interesting

• Native UX, no app UI: The iOS Files app serves as the UI. Drag audio into audio/, read Markdown from transcriptions/. Feels local; works everywhere.
• Beautifully small serverless: Chalice makes Lambda ergonomic, so the whole pipeline stays compact and readable.
• High quality + low cost: Whisper-large-v3 for accuracy; Gemini 2.5 Pro for formatting; end-to-end cost ~ $0.01/min.
• Long-form resilient: Built for long voice notes that turn into essays.

How it works (end-to-end)

1. Upload audio to S3 audio/.
2. S3 event invokes Lambda (Chalice).
3. Whisper (Groq) transcribes to text.
4. Gemini formats to clean, paragraph-structured Markdown.
5. Result saved to S3 transcriptions/filename.ext.txt.
6. Phone notified via ntfy.sh; open in Files immediately.

The core flow in three steps (from the repo)

@app.on_s3_event("audio-to-transcribe1", events=["s3:ObjectCreated:*"], prefix="audio/")
def transcribe_audio(event):
    """

                transcription = groq_client.audio.transcriptions.create(
                    file=(filename, file.read()),
                    model="whisper-large-v3",
                    response_format="verbose_json",
                    language="en",
                    temperature=0.0,
                    prompt="This is a audio recording. Please transcribe accurately with proper punctuation.",
                )

            formatted_text = generate_formatted_transcription(transcription.text)
 
            input_key_without_prefix = event.key.replace("audio/", "")
            output_key = f"transcriptions/{input_key_without_prefix}.txt"
 
            # Save formatted transcription back to S3.
            s3.put_object(
                Bucket=event.bucket,
                Key=output_key,
                Body=formatted_text,
                ContentType="text/plain",
            )

What makes it different

• Files-as-API: Treat S3 like a shared folder synced to iOS/macOS. No custom app, no queue UI, no dashboards.
• Opinionated formatting: Gemini upgrades raw ASR into publishable Markdown—paragraph breaks, grammar, headings—ready to paste.
• Developer ergonomics: chalice for Lambda, uv for fast Python envs. Minimal code, maximum leverage.
• Observability-first: Clean logs, simple notifications (ntfy.sh), and a clear S3 structure for auditing.

Extending it

• Multi-language, diarization, or redaction.
• Per-prefix routing (e.g., different models per folder).
• Signed-link sharebacks or webhooks.
• Cost guardrails (duration caps) and retry/backoff policies.

Security notes

• Keep keys in environment variables.
• Restrict S3 prefixes and IAM permissions.
• Limit allowed input formats at the bucket level.

Style inspiration: Create a Django and React app with auto-generated Django types

• Key innovation: using S3 + Files as the UI, with Chalice to turn S3 events into a clean serverless pipeline.
• Interesting elements: native-feel UX, Whisper+Gemini two-stage quality, tiny operational surface area, and low per-minute cost.