Finite impulse response (FIR) filter is the basic functional component in various signal processing and communication systems. In many practical applications that have stringent requirement on spectrum, long FIR filters are needed to achieve the desired filtering performance. However, because a T-tap FIR filter requires T copies of high-complexity multiplier, the conventional design of long FIR filter consumes a large amount of silicon area and power dissipation. This paper, for the first time, proposes a high-accuracy stochastic computing (SC)-based FIR filter design. By utilizing the simplicity of stochastic arithmetic unit, the proposed stochastic FIR filter achieves significant reduction in hardware complexity as compared to the conventional design. More importantly, this paper proposes a new high-accuracy non-scaled stochastic adder that has significant increase in computation accuracy than the conventional stochastic adder. Built on this new stochastic adder, the proposed stochastic FIR filter achieves much higher accuracy than the existing stochastic FIR filter design, especially for large T cases, thereby unlocking the potentiality for the widespread applications of stochastic FIR filters in practical signal processing systems.