FastSpeech 2

by Microsoft Research / community

Non-autoregressive TTS reference implementation — fast and parallelizable.

TL;DR

Non-autoregressive TTS reference implementation — fast and parallelizable.

Best for embedded TTS where inference latency matters more than absolute quality. Pricing: free (MIT).

Category
Open source
License
—
Stars
—
Last push
—
Pricing
free (MIT)
Platforms
Linux

What it is

FastSpeech 2 from Microsoft Research is a non-autoregressive TTS architecture that dramatically reduced inference latency vs. Tacotron-style models. Still widely cited; ming024's MIT-licensed PyTorch port is the canonical implementation for self-hosters. Consent posture: synthetic-only architecture.

Best for: Embedded TTS where inference latency matters more than absolute quality.
Watch out for: Older architecture (2021) — VITS and XTTS now beat it on naturalness.

Install / use

git clone https://github.com/ming024/FastSpeech2

Features

Speaker diarizationNo
Word-level timestampsNo
Streaming / real-timeNo
Languages supported1
HIPAA eligibleNo

FastSpeech 2 vs Whipscribe

FeatureFastSpeech 2Whipscribe
CategoryOpen sourceTranscription APIs
Pricingfree (MIT)free beta
Speaker diarization—Yes
Word timestamps—Yes
Streaming—No
Languages199
PlatformsLinuxWeb, API, MCP

Alternatives to FastSpeech 2

OpenAI Whisper
OpenAI

The reference open-source multilingual ASR model from OpenAI.

OSS · MIT ★ 98.1k
whisper.cpp
Georgi Gerganov

C/C++ port of Whisper — runs on anything, from a Raspberry Pi to Apple Silicon.

OSS · MIT ★ 48.8k
faster-whisper
SYSTRAN

4× faster than reference Whisper using CTranslate2 — production sweet spot.

OSS · MIT ★ 22.3k

Whipscribe is a managed faster-whisper + whisperX service. If you want transcripts without running infrastructure, paste a URL or drop a file in the form below — you'll have a transcript in seconds.