Blog
What I test and why
Notes on STT and TTS: model benchmarks, provider choices, lessons from shipping voice features.
- 5 min read
Armenian speech to text: what transcription models return
One Armenian recording, clean and noisy: Whisper turbo got 55–77% of words wrong, Deepgram Nova-3 43–55%, two services returned Latin or Arabic script.
Read article → - 4 min read
Groq Whisper 502 service_unavailable on long audio chunks: it's the timestamps
Groq Whisper returned 502 service_unavailable on every chunk of long audio except the first. The cause was OGG timestamp offsets; one ffmpeg flag fixed it.
Read article → - 4 min read
Kyrgyz speech to text: what Whisper and Scribe actually output
We ran 28 Kyrgyz clips from Google's FLEURS set: Whisper via Groq got 86% of words wrong and wrote Kazakh, ElevenLabs Scribe got 10% wrong.
Read article → - 3 min read
ElevenLabs audio tags: what v2 and v3 actually do with them
We ran [laughs] and [whispers] through ElevenLabs Multilingual v2 and v3: v2 reads tags out loud, v3 doesn't. 4 recordings, transcripts, characters billed.
Read article → - 3 min read
How to send text as a voice message in Telegram
Telegram has no built-in text-to-voice button. Send the text to the bot, pick a voice, and get a voice message you can forward to any chat. With sample voices.
Read article → - 6 min read
Speaker diarization: transcription that labels who said what
Hear a three-voice recording, compare the flat transcript with the labelled one, and see measured accuracy on audio where the correct answer is known.
Read article → - 4 min read
How to Transcribe a Telegram Voice Message (2026)
Don't want to listen to a voice message? Forward it to the bot and get text back in seconds. Three ways to transcribe Telegram voice notes, video notes and video.
Read article → - 9 min read
11 STT Models Compared in 2026: Which One to Pick for a Voice Bot
I benchmarked speech-to-text models on voice memos before locking in the STT stack for my Telegram bot. Whisper, Parakeet, ElevenLabs Scribe — what won on accuracy, latency, and privacy.
Read article →